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 2 – Nov 8, 2026

632 stable releases from 157 tracked agents and frameworks in the week of Nov 2, 2026, 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 8, 2026

  • 🌳 See the whole repository: A new Git Graph tab shows local and remote branches and tags, with each commit's details and changed files.
  • 💬 Ask without derailing a chat: In Chat UI, /btw asks a side question in a fork beside the chat while the original keeps running, and /fork branches from the latest response.
  • 🎙️ Dictate from anywhere: Phones can turn speech into text on the device (iPhone on iOS 26 or later, and Android), with the computer as fallback.
  • insert(id, vec, ns, { storedAt?, metadata? }) and insertBatch([{ id, embedding, storedAt?, metadata? }]) store storedAt with every vector. It defaults to Date.now().
  • Hits carry storedAt and metadata, and both survive a close and reopen.
  • search(vec, k, ns, { filter }) passes RVF filter expressions through: { op, field, value } leaves combined with and, or and not.
  • macOS 12 (system SQLite 3.37): observations failed to save with "SQL logic error" on every write. Statements that use RETURNING are now read with .all(), so observations store again
  • CMEM Pro sign-in: since 15.11.0, sign-in stopped at its first step with "Could not start OAuth login". The installer now accepts the setup id that cmem.ai echoes into its sign-in links (#5502.
  • Cursor: the installer writes hooks.json atomically, so an interrupted install or uninstall can no longer leave Cursor with a broken hooks file
  • #413 a85a80a Thanks @riverisbuilding! - Use the built-in OpenFilm agent with the Services account and control the order and visibility of its seven models.
  • #413 a85a80a Thanks @riverisbuilding! - Edit a workspace in place with panel swaps, nested splits, docking and resizing. Undo or redo edits while arranging it, then save the layout or cancel.
  • #413 a85a80a Thanks @riverisbuilding! - Export editing projects with linked originals or full copies, source in/out points, native frame rates, playback speeds and disabled clip states.
HyperFrames
  • CLI: Check replays an unchanged project's last pass and takes --no-craft (070b964d7, #7398).
  • Core: A muted video's picture no longer waits on its own audio track at a cut (fa2e63515, #7412).
  • Producer: Seed random values in every render so re-renders give the same film (eb16510a1, #7410).
  • api: an excluded model skips a gateway leg in every spelling, and the completion gateway sends a leg on its entry's own credentials (b4e2285)
  • api: optional fields on a strict json_schema payload are nullable, and a null there reads as absent (04a534a)
  • discovery: embed the capability index in batches, not one request per capability (3c0fe01)
Sim
  • improvement(search): send repeated account notes and fixed fields once in Search MCP results
  • fix(search): read Search MCP date filters without an offset as UTC
  • @waleedlatif1
  • Make gpt-6-luna crewAI's default model everywhere
  • Read spreadsheet URLs in an agent's input with tools over every row
  • Tag eval usage spans with target and a per-evaluation id
  • Release notes — formatted for people.
  • Changelog — plain Markdown for AI agents and tools.
  • npm package
  • Agent.queue_response() combines background follow-ups with other ready results or a customer turn. Pass after=coroutine to register work before it starts.
  • When tool preambles are encouraged, the agent writes them, and the progress updates that follow, in the customer's language.
  • Agents acknowledge recurring information gaps when different questions cannot be answered, without treating a new question as one already answered.
Oh My Pi
  • Added canReplayRemoteCompactionFrom() to determine whether a model can replay provider-native compaction produced by another model.
  • Fixed GenAI message-content capture configured in omp's own .env files so it continues to apply when those values are no longer exposed through the process environment.
  • Fixed compaction preparation for remote OpenAI compaction that the active model cannot replay, ensuring the content is expanded and remains available for compaction.
No release notes published.
  • esp32agent CLI 0.1.4. Install: see cli/README.md.
  • pi-subagents 0.87.0. A subagent model written as provider/id now always runs on that provider instead of silently falling back to another provider's copy of the model (and billing that account).
Dsh Ops
  • 修复 Windows Niubash 并发调用引发的硬链接计数误报,以稳定文件身份校验执行树,并协调准备中及失败调用的释放;保留零拷贝和 0.5.1 的空间回收机制。 @T-Auto
  • 修复设置页无法保存或清除各模式 Git Bash 执行路径的问题。 @T-Auto
  • 修复离线执行缓存回收遗漏显式及同卷根目录、无法解除只读与拒绝删除封印的问题,并提供明确的 Niubash 故障恢复指引。 @T-Auto
  • fix(dev): dev-piece rebuild clears the cache without scanning every piece folder @kishanprmr
  • fix(engine): raise the disk-pressure threshold to 90% and the idle wi… @MrChaker
  • fix(webhook): stop a failing Return Response from crashing the API process @fabio-kozlowski, @claude
  • Several searches for the job agent: the job agent ran one search. Settings → Job agent now keeps up to three, each with a name, which you can turn off without deleting it or put in another order.
  • The resume, the letter, or both: New application wrote both before you could use either, though many job sites ask for only one.
  • Bold in your bullets: select words in a role's or project's "What you did" and press B (or Ctrl+B) to set them in bold, like a result: "Scaled the system by 40%".
  • No more freezes when one bot hands work to many. When many bot turns waited for memory at once, the server could use all its CPU re-checking the queue and stop answering.
  • Starter runs two tasks at once. Other tasks wait their turn, and your own chat still goes first.
  • Starter keeps five bots awake, including after a restore or a move. Every bot is kept.
  • Restoring a workspace from a backup now correctly preserves the full Git history, even when repository references are stored in packed format
  • When a restore operation fails, the CLI now displays the actual error message instead of a generic failure notice, making it easier to diagnose and resolve issues
  • Import limits for workspace bundles have been raised, allowing large workspaces to be restored or migrated without hitting size restrictions
  • git: fold ls-tree path listings by directory
  • analytics: sum per-project ccusage rows in monthly economics
  • cargo: keep nextest timeouts and retried failures
  • 10327523 job postings. See jobs-latest for the current database. Sources and licences
  • [Departments] Added a schema-only migration option for departments and department-role bindings, allowing existing target records to be preserved.
  • [client-v2] Fixed partial-text search when using string-valued formula fields as labels in v2 association dropdowns and cascaders.
  • [File manager] Fixed blank PDF previews caused by missing JPEG2000 image decoder resources in the v2 file manager.
  • docs: Bring Your Own Cloud guide and the MIT license blog post
  • chore(desktop): bump version to 2.0.5
  • fix(desktop): keep host data cached when a host looks unreachable
  • Let users choose the spoken replies model: Supertonic, ElevenLabs or xAI
  • Show which workflow branch has no answer at a model join
  • Start dev builds with the Dynamic Island and agent dock off
Caveman
  • forward_headers under a providers.<name> block now stops the proxy from starting. Before, it was silently ignored: no headers were forwarded and the only symptom was an upstream 401.
  • caveman-compress refuses a file whose text holds a credential, such as a private key block or a cloud, Git host or model API token.
  • Old recovery handles can now expire. The recovery store used to keep every compressed original forever, so it filled up and compression stopped.
  • The server exits cleanly when the client closes stdin, so no repomix --mcp processes are left behind.
  • The install pulls in about half as many packages (171 to 84), since the v1 SDK brought HTTP transport dependencies such as Express that the stdio-only server never used.
  • Code from function bodies no longer leaks into the output as if it were a declaration.
CodeBurn
  • The macOS menu bar forecasts when a quota window runs out. The forecast uses the pace of the last hour on short windows and the last six hours on long ones, waits for two readings ten minutes…
  • The guard stops loops while they run. An identical tool call repeated 4 times in a row gets a warning and the 8th is denied.
  • Three new Optimize findings for premium spend. one-hour-cache-unused prices one-hour prompt-cache writes that no pause over five minutes ever needed.
Qwen Code
  • fix(channels): keep GitHub inbound recovery armed after a failed read
  • test(core): harden renewal waits and writer seal coverage
  • feat(verify-pr): give the sandboxed round the PR's review threads
  • bash curl -fsSL | bash -s -- --release --version 0.1.2026.11.8
  • Start OpenSRE
  • bash opensre
  • Security: skill the /graphify runbook no longer substitutes the scan path into Python code.
  • Fix: solidity a call on a receiver of a contract or interface type (IERC20 token.
  • Fix: solidity a call attached by a using A for T directive (using SafeCast for uint256.

Nov 7, 2026

  • [PR] fix: revert self-contradictory inclusive billing stamp to split
  • [PR] docs(website): add blog post 'What to Do When the Model Keeps Treating Old Compressed Content as New User Instructions?' | docs(website): add blog post '模型总是把压缩老内容当成用户新指令咋办?' (EN+ZH)
  • [PR] fix: read Responses additional_tools preamble in tool reachability observation
HyperFrames
  • CLI: Cloud uploads send a big file in parts, so a cut connection resends one part (3d7f12acd, #7375).
  • Media Use: The voice credits line names the real source, your recording or any vendor (f5bde9533, #7380).
  • Media Use: Your own voice recording is a voice line, and speed applies to it (14e598e5a, #7377).
AgentOS
  • language: a transcript translator on generateText and a per-participant LiveKit translation output (1a81996)
  • hearing: a send or a replay from inside encode waits for the line being encoded, as a write does (8e866e6)
  • hearing: a translation write whose encode throws rejects its promise instead of throwing at the caller (5fd9ddb)
  • Host-only settings: Repository .pr_agent.toml files can no longer override GitHub pacing and retry settings…
  • Comment arguments: config.ai_timeout and pr_code_suggestions.max_number_of_calls can no longer be set through command comments.
  • Removed no-op settings: pr_code_suggestions.approve_pr_on_self_review and pr_code_suggestions.fold_suggestions_on_self_review are removed from the shipped configuration.
  • AWS Bedrock: Fixed every request failing on a Bedrock deployment with api_base: "", which answered 500 or 400
  • MCP: Fixed tool results over 1 MiB failing with SSE stream ended without a response, broken since v1.103.0
  • Guardrails: Fixed model-level post_call guardrails checking each response twice and calling the moderation provider twice
OpenMausBot
  • One way to connect. In the app, + → Connect to a server takes the link your server prints, or its address and a pairing code. A plain address works on your own network (Wi-Fi, Tailscale, .local).
  • One pairing code. Servers and the web /pair page use the same 12-character code, and a code of the wrong kind says so.
  • Commands that work when pasted. The server prints one link per way in, checks that the address really reaches it, and prints every command the way you started it (npx, a global install or Docker).
  • Managed-agent sessions can hold multiple child questions at once, answer them in any order, and keep drafts and errors separated per child.
  • Web Shell adds a task Q&A heatmap and an Ask for details side task that explains pending Ask User Question options without answering.
  • GitHub channel editor adds Repository Filter allowlists and blocklists, while bound PR delivery and plain comment delivery become more reliable.
  • 10327476 job postings. See jobs-latest for the current database. Sources and licences
Claude Mem
  • Kept by default: screenshots and images your agent reads or makes are now kept and linked to the observations from that turn, without sending them to an AI model. Image inference stays off.
  • "Keep images" switch: the memory viewer's Settings has a new Images section. Turning Keep images off stops keeping new images; images already kept stay. A "false" you saved yourself is respected.
  • Gallery in My Work: request cards show an image count and the first four thumbnails; the request drawer shows every linked image; any thumbnail opens a full-size gallery with keyboard navigation.
AI SDK
  • 883fcde: remove transcription word timings from telemetry when recordOutputs is false
  • 6d21e8e: Sanitize provider request URLs before text generation results reach telemetry integrations and tracing subscribers, preventing endpoint credentials from being exposed.
  • Updated dependencies [6400cf8]
  • 更新前请注意:这一版的运行时有更新,所以还是下整包。 内置引擎仍是 DSH 0.2.2-rc.1,变的是我们给它打的补丁。升级后 Linux 环境里只保留 /root,用 apt 装的软件和放在 /usr/local、/opt 等目录的文件需要重新装。对话和设置不受影响。
  • 这一版没有在真机上验证过。 App 内的视频播放和存视频到相册只经过自动测试和编译检查,没在手机上放过。遇到问题请开 issue 反馈。
  • 新对话首页:建议换成横排小卡片,下面一行小技巧会自动轮换,点一下换下一条。首页就能选这次对话的模型、Agent 预设和权限,第一条消息发出前生效;设置失败时消息不会发出,留在输入框里。
  • alerts: let a card ask for an answer, a pick, an OK, or nothing (12bc909)
  • briefing: fold old Needs you cards and group a request's approvals (5f87e70)
  • briefing: show Home as Needs you and Done for you, each item once (8194ba5)
Esp32 Iphone Agent
  • Hear the phone (optional): the board forwards AirPlay audio, the app transcribes it on the iPhone and the agent knows what it heard.
  • Long tasks: no 40-round limit, and any stopped or failed run can Continue from where it was.
  • A more accurate pointer and steadier video (stream drops 41 → 0 in a Duolingo lesson).
Reasonix
  • CLI archives built from the Studio release studio-v2.58.0.
  • agent-framework-core, agent-framework-foundry-hosting: Support callable MCP discovery approval policies and honor Foundry Toolbox approval metadata, including tool-name and read-only filters…
  • agent-framework, agent-framework-a2a, agent-framework-ag-ui, agent-framework-anthropic, agent-framework-azure-ai-search, agent-framework-azure-contentunderstanding, agent-framework-azure-cosmos…
  • agent-framework-core, agent-framework-declarative: [BREAKING] Bound MCP tool and prompt discovery to 100 pages and 1,000 items per catalog, default framework-owned request timeouts to 30 seconds…
  • Fix races in implicit state machines
  • Open files that a chat message names on mobile
  • Keep no server rail space in a chat window
  • Godot 4.2+
  • Active Ziva account
  • Control characters in an echoed PLAN_ID, PWF_PLAN_ROOT or target print as ? on both routes, so a newline or escape sequence cannot print a fake status line (PR #362, #358).
  • attest-plan reports the multiple-plans refusal once, exits 1 and attests nothing, without the extra No task_plan.md found line (PR #362, #358).
  • attest-plan.ps1 gains -Target root|<plan-id> and one-line refusals that name a rejected PLAN_ID, PWF_PLAN_ROOT or -Target.
Orca
  • Reduce repeated work during host-status updates
  • Reduce process launches when reading Linux process start times
  • Dim the chat message rail until hover, focus or scroll
  • Real devices can now reach private backends through your LT tunnel. Pass --tunnel or --tunnel-name <name> to run, testmd run or testrun run.
  • On web, a type step can ask for a delay, so fields that ignore a normal fill are typed key by key. The delay carries through replay, upload and code export.
  • Android vision taps now map to full-display pixels.
  • Added each subagent's model next to its name in the subagent rows above the prompt and in /agents (hidden on narrow terminals), with the full model and effort in a subagent's detail view, and a…
  • Added the PLUGIN_ROOT and PLUGIN_DATA environment variables to hook commands that come from a plugin, pointing to the plugin's folder and its persistent data folder, so plugin hooks can use files…
  • Changed the error shown when Vertex AI reports that the selected model does not exist or that your Google Cloud project cannot use it, which now reads The requested model was not found, or your…
  • Gateway fixes: a dropped Postgres connection at the gateway boundary now returns a retryable 503; a non-streaming reasoning tool turn cut by max_tokens settles as length instead of a 502.
  • Mika-v0.1.38-arm64.apk:近几年的安卓手机基本都选这个,更小。
  • Mika-v0.1.38.apk:通用包,所有手机都能装。不确定就选它。
  • 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
  • Codex returns control when an automatic memory request times out.
  • Automatic checks have more time to finish when a response is slow.
  • Work passed between several agents stays with its current owner.
  • Memory and CPU: a long pass through refresh and UI work. In synthetic tests, first opening Usage & Spend uses about half the CPU and far less memory (about 680 to 118 MiB), plugin refreshes use…
  • Cost history: Claude and Codex history stay complete through symlinked directories, Claude and Vertex avoid most of the midnight rescan, AWS Bedrock keeps zero-cost days, and local day boundaries…
  • Providers: new SiliconFlow balances and Tripo credits. Cursor reads current encrypted app sessions, TypeSafe billing is restored, Kiro recovers API usage from a CLI token, Synthetic shows weekly…
  • Added support for disabling the per-call timeout.
  • Fixed interpreter shutdown blocking in CodeInterpreterMiddleware.__del__.
  • Fixed resolution of nested schema $refs in PTC signatures.
  • Credit codes: redeem a code at /redeem for credits, and a /redeem?code= link survives sign-up. Admins create codes with an expiry and a redemption limit.
  • Goals run on the API: goal state lives on the server, and only a verifier session can mark a goal done. Goals are in the command palette.
  • Proactive Spaces: a watcher reads a Space's event feed and speaks up only when it matters. A Space's agents are its router's team; the router does quick work itself and links results as files.
  • Signed Caveman runtime companions for macOS, Linux, and Windows. The binaries are licensed under the Apache License 2.0 (LICENSE, NOTICE); see LICENSING.md and the attached third-party notices.
  • feat(media): add Higgsfield to video and image generators, allow typed models
  • feat(higgsfield): add Higgsfield integration
  • feat(search): add Hex project search and notebook reads
  • Add functionality to read spreadsheets for agents in crewai-files.
  • Scaffold and load standalone agent projects.
  • Update RagTool description for better parsing by Mintlify.
  • Fixed Mistral discovery listing embedding, OCR, moderation, transcription, and speech models as chat models.
  • A failed refresh no longer replaces a cached catalog whose models carry request headers with an empty snapshot; the next launch refetches it instead
  • Fixed transcript output freezing after provider restarts or switching sessions mid-stream
  • Opening Settings → Local agents no longer stops Bot settings and other DS Bot dialogs from opening until the window reloads.
  • The Docker image runs DeepSeek Harness 0.2.2-rc.1.
  • df1cb33d0d32699c8d6a2e5b3ea7e5f8990d86d5 .Net: Bump package version to 1.82.0
  • d08a482fab9c6980fa50de5eee79895ba702069d Python: Prepare 1.47.0 release and consolidate dependency updates [ #14708, #14596, #14597, #14598, #14599, #14600, #14627 ]
  • 28c0d60222b128efc5a41d825ea74e8836231e9c .Net: Harden structured data mutation boundaries
  • Execution environments and sandboxes
  • BaseEnvironment, LocalEnvironment, EnvironmentToolset, DaytonaEnvironment, and E2BEnvironment are now stable. (c54b253, 799fc94)
  • BaseEnvironment.read_file_stream streams files in chunks across LocalEnvironment, E2BEnvironment, and DaytonaEnvironment so ReadFileTool reads line ranges without loading the whole file into memory.
  • bash curl -fsSL | bash -s -- --release --version 0.1.2026.11.7
  • Start OpenSRE
  • bash opensre
Phoenix
  • Updated dependencies [6e6e101]
  • @arizeai/phoenix-client@7.18.2

Nov 6, 2026

  • Shared MCP Connections: Turn on pooling for an MCP server to share one connection across projects with the same setup, so you run fewer processes and start faster.
  • GUI Plugin Themes and New Slots: GUI plugins can now add color themes, render or add actions to code blocks, and use native settings and menu components.
  • Mobile Composer Dock: On phones and tablets, composer actions now sit in a touch dock above the keyboard with larger buttons.
Phoenix
  • otel, client, config: parse .env.phoenix the same way in Python and TypeScript (67cc379), closes
  • otel: pick the OTLP protocol once (0f5024c)
  • otel: show the header name in the URL-encoding warning (30949b8)
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
  • Portable narrated research films. Export any session as a portable, narrated research film with terminal playback and voice narration: narration assets download from the app's built-in CDN first…
  • Literature tools in the CLI and MCP. Search and manage the literature library from the command line and from MCP clients, with read-only browsing plus write tools for keeping references up to date.
  • Resumable batch paper overviews. Batch overview generation picks up where it left off after interruptions, without regenerating results that already completed.
OpenMausBot
  • Phone connection: when you set a phone to "Wi-Fi and Tailscale only" (or back), the phone hears about it straight away inside its encrypted connection, so it shows the right status instead of…
  • Email sends with any provider, and says why when it can't.
  • OpenCode: one thread per folder in a shared process, fixing the mid-turn "capability has expired" error; the model list refreshes on its own.
  • The Pipelines experiment is removed. Pipeline boards, stage settings, case automation, review queues, the pipelines CLI commands, the pipelines:write permission, the enablePipelines setting, and…
  • Run failures no longer reach Sentry by default. A backend DSN alone no longer reports agent run failures.
  • The run history list is bounded. GET /api/companies/{companyId}/heartbeat-runs returns 200 rows by default and at most 1,000; clients that relied on an unbounded array must pass an explicit limit.
AI SDK
  • ba3743f: Preserve toolChoice: 'none' when SDK tools are omitted or empty so providers can disable native tools configured through provider options.
  • Updated dependencies [fbb8489]
  • Updated dependencies [26b7351]
Claude Mem
  • Track the nine actual setup screens: Scan, Reveal, Choose memory, Review savings, Checkout, Finish setup, Import history, Launch, and Workspace.
  • Count each rendered screen visit separately. Returning to a screen creates a new visit; failed hand-offs do not become successful forward transitions.
  • Preserve the real setup branches. At Home and existing Pro configurations skip checkout, history import returns to Finish setup, and reduced-motion setup skips the launch animation.
  • OpenAI: Fixed openai/ deployments with an empty api_key sending OPENAI_ADMIN_KEY as the bearer token
  • AWS Bedrock: Fixed Converse service_tier so unsupported tiers answer 400 and reserved, auto and standard_only are honored
  • AWS Bedrock: Fixed Invoke dropping the per-turn-control-2026-07-01 beta on Claude 4.5 and newer
  • Foundry: Add the GA computer tool through FoundryAITool.CreateComputerTool() and carry computer calls and results through Foundry hosting, including batched actions.
  • Foundry hosting: Add FoundryToolboxOptions.TokenScope for cloud-specific toolbox authentication audiences, retaining as the default
  • A2A hosting: [BREAKING] Limit built-in server registration to message-based responses while reliable task resumption across hosts is deferred
Loki Mode
  • Never-stop plus one Opus consult (LOKI_E10_OPUS_CONSULT, default on, =0 opts out): a run that would stop on an ambiguous or conflicting spec implements its best reading, discloses the choice and…
  • Cost (default on, each with an =0 opt-out): the claim re-ask that never edited is skipped, the coverage turn is capped at 10 turns (LOKI_E10_COVERAGE_MAX_TURNS), and a fix round is skipped when…
  • The run-state guard no longer blocks the files a stage's own brief tells it to write (plan output and scope, project-model answer, spec criteria); run record files stay denied even if listed.
  • #370 4287d3c Thanks @riverisbuilding! - The assets drawer has a library: its categories along the top (Media, Voices, Sounds, Music, Text, Stickers, Filters, Effects, Transitions, Templates) and…
  • #370 4287d3c Thanks @riverisbuilding! - Canvas: moving a card moves the card, never the board — the board only glides to show a card's panel after a plain click.
  • #370 4287d3c Thanks @riverisbuilding! - Canvas: a project can have several canvases (new, rename, duplicate, delete into the trash, open in a new window), and each can be seen as a storyboard of…
  • Remove the Fern config and the frozen OpenAPI copy (SKY-20668) (1620bdcd7)
  • Store org and workflow agent skills in a versioned table (SKY-20416) (f3942b1ce)
  • Give a tokenless Copilot Accept an outcome the client can read back (SKY-17870)
Sim
  • fix(billing): settle hosted fal generations to their exact cost
  • fix(mothership): answer a Chat workflow tool paused for a person with its resume link
  • fix(billing): refuse Stripe checkout for AWS Marketplace-billed customers
  • AgentRun observation CLI 0.1.10 with the Node launcher @parcha/agentrun-cli 0.1.1.
  • Connect selected Claude Code and Codex sessions to an AgentRun workspace using device authentication or an explicitly configured enrollment credential. Observation reports activity.
  • Verified platform: Linux x64 with glibc, Node.js 20 or newer. The launcher provisions Python 3.12, verifies release checksums, and installs a persistent runtime outside the npm cache.
  • bb desktop 0.49.0
  • Defer allocator reclamation until disposed session data is released on opted-in Linux/glibc hosts, and preserve requests made during trim cooldown.
  • Sandbox deny rules support glob patterns, including hidden files and home-relative patterns, without changing literal grants.
  • Session-summary failure diagnostics omit private parent directories.
No release notes published.
GitHub Copilot CLI
  • Rex terminals render inline images using Kitty direct placements.
  • Add /code command to open the current session in VS Code
  • Add a Fusion tier setting to copilot config.
LoopForge
  • fix(web,chatbot): copy GitHub device code on HTTP; rename remaining AgentFlow labels to LoopForge
  • fix: show LoopForge instead of AgentFlow in user-visible text
  • docs: keep bash scripts/install.sh as the quick start alongside ./start.sh
  • 更新前请注意:这一版的运行时有更新,所以还是下整包。 内置引擎仍是 DSH 0.2.2-rc.1,变的是我们给它打的补丁。升级后 Linux 环境里只保留 /root,用 apt 装的软件和放在 /usr/local、/opt 等目录的文件需要重新装。对话和设置不受影响。
  • 这一版没有在真机上验证过。 目前在 x86_64 模拟器上用测试模型走过从 1.0.11 升级:运行时完成替换,对话和定时任务都在,/root 保留、/usr/local 清掉,升级后发消息正常。伙伴小屋只经过自动测试,没在模拟器上打开过。 遇到问题请开 issue 反馈。
  • 伙伴小屋(新):从「切换主机」面板进入的 Live2D 伙伴页面,可以摸摸她、和她聊天。轻量模式由手机直接调模型,不用启动引擎;完整模式在当前主机上开一个普通会话,能用主机的工具,审批照常。她会记得你写下的事;可以用手机自带的朗读或你自己配置的语音服务说话,也能用你自己的录音克隆音色,录音只发给你选的那个服务。Live2D 运行库和角色模型不随 App 提供,需要自己导入素材包。
Metaharness
  • createRvfBackend({ dimensions, metric?, dir?, m?, efConstruction? }). dir is where the store persists. Without it, the backend uses a temporary directory and removes it on close().
  • null is returned only when @ruvector/rvf is not installed (or has no RvfDatabase). Every other failure throws: a missing native binding, a held lock, or a dimension mismatch.
  • Namespaces are checked against [A-Za-z0-9._-]{1,128} and may not start with ., because they become file names.
  • bash pip install claude-agent-sdk==0.2.179
AI
  • #2047 2641779 - Raise the Angular peer range to >=22.0.0 <23. The package is built with Angular 22 and TypeScript 6.0.2.
  • Safer ingestion: raw database connection strings and unsafe server/file paths are refused by default, reducing risk of accidental credential or server exposure.
  • Search and recall correctness: typed/scoped queries (queries targeted at a specific document type or scope) now return correct, LLM-safe results and "feeling lucky" picks ignore disabled data types.
  • Retrieval performance: query scoring is much faster for PGVector, Turso and LanceDB backends by batching and parallelizing scoring work.
  • Changed auto mode with telemetry or feature flags off, including behind LLM gateways: calls the server can't check are asked about or denied, with no fallback to the built-in classifier.
  • Changed auto mode behind an LLM gateway: CLAUDE_CODE_AUTO_MODE_SERVER=1 turns on the server-side check even with CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS set.
  • Added support for MCP tasks in the terminal: when an MCP server runs a tool call as a task, Claude Code moves it to the background right away and tells Claude the result when it finishes
  • Adds PUT /skills/activation and setUserSkillsActivation to enable or disable 1-100 user Skills in one request, with saved preferences and propagation status.
  • Prepares Skill installation for larger archives by adding disk-based ZIP handling and recovery-backup protection to the installer.
  • Adds daemon API routes for installing large Skill ZIPs by upload or host file path and querying installation receipts after failures.
  • Fixed conversation text becoming blurred when opening or dismissing the keyboard on iOS 26+
  • Fix TLS certificate verification, sandboxed shell execution and workspace permissions for queued inputs.
  • Keep idle maintenance from blocking active work or incorrectly refusing an update. Improve retry, reconnection and recovery behavior, and preserve native provider state through compaction.
  • Add a file-context graph with file previews, local Markdown links and recent context snapshots. Switching topics keeps each conversation's context separate.
  • This release introduces native Generative UI for Angular, adds analytics context forwarding for Intelligence runs, and fixes several reliability issues around Intelligence run stops and A2UI…
  • 235518c: Include the tenant id in the metadata of chat messages ingested into memory.
  • Bots can use the apps on your computer, as far as you allow: turn it on in Settings → Computer use, give each Bot a level and its apps, and take over or hand back at any time.
  • Talk to your Bots from WeChat, Feishu, DingTalk, WeCom, QQ, Telegram, Discord and Slack, with pairing, group chats, approval cards and replies (Settings → IM).
  • Connectors: connect services from a catalog with one sign-in, add your own or import them from other apps, connect QQ Mail, 163, Gmail and other mailboxes, and choose which Bots use each connector.
  • pi-subagents 0.86.0. A session no longer loses its subagent tool after a compaction when pi-subagents was reloaded or updated earlier in that session.
  • fix(renderer): export live rendered content through shared markdown conversion
  • fix(markdown): export composed Shadow DOM content #2376
  • perf(markdown): export live DOM with lazy projection caches #2402
Ruflo
  • @ruvector/attention 2.2.3.
  • claude-flow 3.72.1 declared @ruvector/attention ^0.1.3, and ruvector 0.3.12 also depended on ^0.1.3.
  • 3.72.2 moves to ruvector ^0.3.13, which depends on attention ^2.2.3, in the CLI, claude-flow and ruflo. claude-flow also declares attention ^2.2.3 directly.
  • Keep an old agent reply out of the end of the chat
  • Add Higgsfield to the Apps marketplace
  • Find MCP server commands on the PATH when the login shell is slow or fails
  • feat(automations): explain each automation with a Mermaid diagram
  • fix(chat): place resumed sessions after UTC conversation history
  • fix: keep generated conversation titles visible during startup handoff
  • Rewritten in Go for lighter deployment: Docker images are approximately 10× smaller, and startup is approximately 10× faster.
  • Approximately 100× greater concurrency, with substantial improvements in query latency, memory usage, and document parsing speed.
  • All-new Agentic Chat: Enables multi-step reasoning and answers to complex questions through iterative retrieval, tool calls, and evidence sufficiency assessment.
  • Cloud on for everyone, faster Settings, share gate, file attachments
  • Dictate into the composer with native macOS speech recognition
  • Add Hermes Agent harness with native ACP chat
  • Andee娘 replaces 大肥鱼. The same slot, but the idea now lands on our own character instead of
  • If you had picked her, nothing happens. The saved id reads as the new one — you keep the look, the
  • She keeps the same voice and the same temper (sweet, soft, a little shy, earnest about the work).
  • Preserve LangGraph Store key escaping, empty-index handling and backend write/list error handling
  • Preserve Unicode archive query constraints in text search
  • Review: agent questions and permission asks no longer block a session. After a short live wait they park in the new Review page, and the answer reaches the session as a new message.
  • Spaces: Slack-style spaces where every message is a thread, with a router that hands work to agents and agents that post into spaces (flags Spaces-Router, Spaces-ChatInput).
  • Live co-editing of Artifact docs: people and agents type together, with presence, carets and pointers. Artifact cards show open comment threads.
AgentOS
  • api: carry inclusiveInputTokens through generateObject's usage across attempts (25e37ba)
  • embed: embedText's own requests send their reservation's operation id (165aab8)
  • embed: Gemini's embeddings take the header name embedText read before its reservation (92ab5a7)
  • commit: 9d62b213c960c63daa46996de8a606965f2f2299
  • source tag: beta2026-11-06_18-14-41
  • regions: cn global
  • 3faf996d92cb666bb21050957e1b30043614b542 MUL-4066: feat(skills): skill version history and restore
  • f02a6aa60e572d3038b06328b565eaeb9b3a4fd6 MUL-7563 perf(chat): virtualize session history lists
  • 16ce2df18c325cce1b3760704169373b4f741086 MUL-7782 feat(plugin-api): task transcript endpoint GET /v1/tasks/{task_id}/messages
  • Blog: ThinkRail 0.3.1 release post
  • feat(ask-user-question): submit from the last question instead of a Review & submit page
  • fix(web): let an active IME own Enter, Escape and arrows in every text field
  • [OPIK-8934] [SDK] fix: Prompt() and ChatPrompt() use the default project
  • [OPIK-9097] [BE] feat: map ClickHouse TOO_MANY_BYTES to a clean 400 and cap free-form SQL bytes read
  • chore(deps): Lock file maintenance
  • #4728 7b906ea Thanks @sailist! - Fix duplicate tool declaration errors in select_tools when MCP tool definitions change.
  • #4731 ef10107 Thanks @liruifengv! - web: Fix steered messages being incorrectly queued while a session is running.
  • #4731 ef10107 Thanks @liruifengv! - web: Fix project creation when accessing the Web UI over HTTP.
  • Update notices. Once a day, Poki checks GitHub for the latest release. When a newer version is available, poki_mcp_status and poki status display the version number, a one-line summary, the…
  • poki update. In a source checkout, this command fetches the latest code and rebuilds it using ./bin/setup. It only runs when you are on main with a clean working tree.
  • Restart prompt. When the code has been updated but the service is still running the old version, Poki tells you to restart it.
  • Windows: download AutoSwarm-Desk-Setup.exe and run it. It installs for the current user, includes Python, and updates itself.
  • macOS: download AutoSwarm-Desk-mac-arm64.dmg for Apple silicon (M1 and later) or AutoSwarm-Desk-mac-x64.dmg for Intel Macs, open it and drag AutoSwarm Desk to Applications. Python is included.
  • Windows: 下载 AutoSwarm-Desk-Setup.exe 并运行。安装到当前用户,自带 Python,会自动更新。 免安装:解压 AutoSwarm-Desk-0.1.0-win-x64.zip,运行 AutoSwarm Desk.exe。
  • 📱 Much more chat on the phone: Rewind to an earlier message, continue a stopped reply, pin the agent's plan above the composer, paste images and long text, manage a chat's goal and see context…
  • 📎 Add anything to chat: Send a selection from a file, a Markdown preview or a terminal to any agent in the workspace, or start a new agent with it.
  • 🔁 Chats pick up after a restart: New installs resume working chats automatically after Orca quits or updates, and resuming is more reliable for everyone who has it on.
  • Added skills from the cloud plugins you connected with /plugin to --remote-agent sessions, so they now appear in the / menu and can be typed at the start of a message, /skills lists them grouped…
  • Changed the default for macOS Option-key shortcuts such as alt+m, which now work out of the box without setting Option to act as Alt in your terminal, where previously pressing them did nothing or…
  • Changed --dangerously-skip-permissions so commands on the system command denylist are blocked instead of approved automatically, while every other command still runs without asking.
  • Added Agent.queuedMessageGate to filter queued messages before delivery and observe claimed messages returned undelivered.
  • Added wireToolsFingerprint(), returning a hash of the tool specs sent with a request.
  • Added opt-in standard Responses WebSockets for catalog-supported models, with incremental turns and automatic HTTP/SSE fallback; enable through OMP settings
  • Add experimental conversational Agent sessions.
  • Ask for a missing API key when crewai run starts.
  • Record the error class when crewai run cannot load a project.
  • 10251211 job postings. See jobs-latest for the current database. Sources and licences
HyperFrames
  • CLI: A font missing from this computer no longer stops a cloud render, and is named (b56e2c036, #7287).
  • Core: A render lands where preview does when a set interrupts a tween (6abc62a17, #7283).
  • Windows 默认 Shell 改为随主包携带的完整官方 Niubash v1.5.0 载荷:逐条校验归档、源树与每次调用的执行副本,缺少载荷时不隐式回退到环境 Bash,并提供托管安装、状态、选择与卸载命令。 @T-Auto
  • 内置极简、标准、PTC 与创造预设默认隐藏宿主 PowerShell,仅在 Bash 替代工具可调用、总闸允许且未开启 dsh-win-pwsh7 联动时生效;已保存的显式值与自定义预设保持原意。 @T-Auto
  • 插件详情页改用 locale 的 meta.detail 展示专门编写的介绍文档,可含图片,不再直接使用仓库 README。 @T-Auto
  • 适配 DSH 0.2.2-rc.1 — npm 上的 DSH latest 已是 0.2.2-rc.1,而 v0.18.0 的精确 peer 只到 0.2.1-rc.5,升级 DSH 后 TUI 会被判 incompatible 并 skip。本版把 0.2.2-rc.1 加入验证线与 peer 范围,0.2.1-rc.5 及其之前的验证线继续保留;模型选择改用上游的会话级…
  • 一次 /update 对齐 TUI、dsh 引擎与全局启动器 — 以前 /update 只换 profile 里的 TUI,全局 dsh 与启动器留在原地,可能出现新 TUI 的 peer 范围装不上当前引擎。现在预检为三层解析同一份计划,按 TUI → 引擎 → 启动器串行执行:引擎只选目标 TUI peer 范围内已发布的最高版本(永不降级、不选 alpha 线),装完用…
  • 子代理通道整体失败会明说 — 同一 provider 的已结算运行至少失败 2 次且失败过半时,子代理列表上方出现通道降级提示,并建议切换 provider;只有可切换的 claude-code / codex 通道会提示,内建 subagent、workflow 与 teammate 行不误报。
  • 工作台增加云主机页:授权后查看腾讯云 CVM,并由常驻运维专家巡检和变更
  • 后台标签页的定时提醒改为系统通知
  • 启动时同步新的预置模型,腾讯云通道支持 HY4
  • Open the agent command center with codex agents --no-daemon. The command center and session picker now mark unread results.
  • Switch between connected backends from the terminal session picker and keep your drafts when switching.
  • Preview local image attachments while composing on supported terminals, and open viewed local images from clickable transcript paths.
  • fix(bitbucket): scope repo code search with the repo: qualifier (fe2908738) — Dhruvin Krutarthkumar Dave
  • fix(vercel): read runtime logs from the project endpoint and report failures (c0496c5d7) — Dhruvin Krutarthkumar Dave
  • docs: document Pipedream and scheduled email delivery; remove stale Yandex Cloud file (54dad0684) — Awokoya Olawale Davidson
  • Fix Responses WebSocket captures expiring while waiting for HTTP relay metadata. Native response capture, settlement eligibility, destination acknowledgement, and capacity limits remain enforced.

Nov 5, 2026

  • A desktop app for Windows and Linux: download it from this release, under Assets.
  • Your own sections: Profile → + Add a section makes a section of your own, such as Awards, Courses, Talks or Interests.
  • Sections in your order: on the Profile page, drag any section to where you want it, or use the arrows. Skills can come first, for example.
  • The cosign claims were validated
  • The signatures were verified against the specified public key
  • ci: run the OSV scan and unit tests on stable, rc and lts branches [backport lts/1.108.x]
  • fix: patch release 1.12.10 image vulnerabilities
  • fix: patch bundled npm glob and minimatch for 1.12.10
Loki Mode
  • A Go test file behind a build tag the run does not set ("[no tests to run]" with no test executed) is read as a zero run, the same as on the original commit, instead of capping the run below VERIFIED.
  • A failure on the original commit that raises the exact error type and message the issue reports as current behavior counts as fail-then-pass proof, when the same test asserts the call result…
  • When the only failures are ones the original commit already had, the run no longer stops on SPEC_CONFLICT asking which behavior to keep.
OpenMausBot
  • The whole app is now in Brazilian Portuguese.
  • Japanese gets the 836 strings it was missing, and one set of terms throughout.
  • On a Team Cloud, each person keeps their own language in shared rooms.
OpenHands
  • Temporary evidence assets for OpenHands/OpenHands#19394. Browser click-through of automation setup dead ends and decoded PostHog automation_setup_blocked payloads.
  • <!-- macos-install-note --> ### macOS install note
  • The app isn't signed with an Apple Developer ID yet, so macOS may call a downloaded build "damaged" — it's Gatekeeper quarantine, not a bad file.
Deep Agents
  • Size the compaction input budget to match the selected model.
  • Preserve cut-off reply handling when ModelRefusalMiddleware raises an error.
  • Defer in-app uv installations on Windows.
  • Your Desktop workspace in a browser tab — The new Webapp serves the full Desktop workspace from your own gateway: the same chats, Group Chats, plugins and settings, in any browser on your network…
  • One gateway owns every conversation — Before, the CLI, TUI, Desktop, editors (ACP), messaging bots and cron could each run their own copy of the agent against the same chat, which is how messages…
  • Pictures in the terminal — When you paste an image, or a tool takes a screenshot, the CLI and TUI now draw it right in the chat instead of printing a file path: real pixels in kitty and Ghostty…
Esp32 Iphone Agent
  • esp32agent CLI 0.1.2. Install: see cli/README.md.
  • Restore a pull request you marked Not mine. In Reviews, a teammate's pull request in Set aside now offers Restore beside Re-run review.
  • PR actions. prFlows.prActions in config.json adds your own commands as buttons on a pull request in My PRs, next to Keep mergeable and Merge when ready, limited to the repositories in repos when…
  • PR action reports. A PR action that hands work to something slower, such as an agent that merges later, can report back when it finishes: the command, or any process it leaves behind, runs…
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
HyperFrames
  • Producer: Render long videos with windowed source frames behind a flag (968f97fe9, #7242).
  • Producer: A capture segment extracts the frames it shows and deletes them (599eaad2b, #7241).
  • CLI: Check --budget stops the whole check in time and reports what it found (b7d72c947, #7245).
  • The worker waits for a free slot instead of failing when the pool of 10 is full, and releases an idle slot when other sessions are waiting.
  • Hung context-reader requests now time out, so a stuck read no longer stops observations.
  • Failure tracking restarts when the kind of failure changes.
  • Zoom AI: Zoom Scribe STT with live streaming and per-utterance recognition.
  • Kalpa: kalpa-tts-1 streaming TTS in English and Hindi.
  • Rime: STT and Prism speech-to-speech through the standard realtime interface; both require Python 3.11+, and STT requires the stt extra.
Headroom
  • ccr: keep the in-memory order queue from evicting re-stored entries (c6f5bf9)
  • memory: read the relation key the memory_save schema advertises (7592e68)
  • memory: replace the HNSW vector when a memory is re-indexed (ee638f9)
  • Release Notes
  • Breaking Changes
  • @AmanDeka made their first contribution in
Qwen Code
  • refactor(managed-agent): Share the session event delivery loop
  • ci: route the classify_pr gate to ECS for every pull request
  • fix(core): delete the OAuth refresh lock only when owned
  • 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.
  • Linear runs link their issue: the branch, PR title and body carry the key (Fixes KEY on the final slice, Part of KEY before it), park, env-failure and slice-plan comments go to the Linear issue…
  • A queued issue whose prerequisite has an open factory PR now builds on it and opens a stacked PR against that branch, restacks onto main when the prerequisite merges or onto its new head when it…
  • Feedback rounds: review rework reads all trusted PR feedback, re-plans on scope changes, replies to each item, resolves fixed threads and re-requests review; factory respond <pr> starts one by hand
  • feat(sessions): Add chronological session list view
  • feat(telemetry): report AI gateway key count from self-hosted instances
  • feat(topics): add Topics summaries, assignments and rules to the public API, MCP and in-app agent
  • d61ee823196151e9c282d497a78ec66a800ea91f feat: add Casdoor Cloud links to the console and a global setting to hide them
  • 04a36b79e3d8f19cc70b68d617edabccd9a9cc83 feat: add initAdminPassword to Helm and docker-compose
  • f3f904852c62de0d523efe25b1eab8c6207438a7 feat: ask again on the prompt page when the display name is not a real name
  • permissions safe marks commands safe for yolo
  • permissions block blocks commands completely
  • permissions unblock unblocks restricted commands
  • opencode-pty 0.5.0 (protocol 8) from packages/process-daemon, published to npm as @opencode/pty@0.5.0.
  • ⚠️ Unsupported release. Upgrade to the latest release
  • See CHANGELOG.md for the list of changes in this release.
  • C#: preprocessor directives and #if branches that derail tree-sitter are blanked before parsing (byte-length preserving), so a directive no longer closes a class early and drops later members or…
  • Solidity: library calls (Lib.f(), including aliased imports) and emit/revert/require naming a declared event or error now link to the member they name, gated to a unique visible type and member.
  • YAML: a << merge sequence gives the earlier anchor precedence for a duplicated key, matching the spec and PyYAML.
  • pi-subagents 0.85.0 and pi-web-access 0.47.0. Workflow scripts that launch parallel researchers now run in a locked-down process, a research prompt with long runs of separators or git options no…
  • Poki MCP: a local MCP server that lets your agent read and operate a real iPhone. Connect it to Claude Code, Codex, or another MCP client and use your own subscription. No extra model account needed.
  • Poki Agent: send a task from iMessage, Telegram, or the terminal. It runs the task on your iPhone and learns reusable routes.
AI Memory
  • Projects split by a git remote that arrived after capture began. 3.4.1's repair renamed the two projects but left the repository identity on the archived one, so capture and later launches could…
  • OMP and Pi users: regenerate the extension after upgrading (ai-memory upgrade, or ai-memory install-hooks --agent omp --apply / --agent pi) so ranged reads and @/file:// paths of ignore_paths…
  • Restore no longer undoes a session purge. Restoring a purged session's summary is now refused; git history and backups remain the recovery path.
  • 手机 App 后台提醒(#1780):离开 App 后的审批、完成和出错提醒不再依赖发布者的推送密钥,也不需要自建服务器。iPhone 由 Expo 推送服务转发系统通知,电脑需要能访问 exp.host;Android 由通知栏里常驻的「正在值守」向电脑拉取,不依赖 Google 服务,部分国产手机需要在 App 的「设置 → 通知」里允许后台运行。手机设置里会显示后台连接和电池优化状态。
  • 手机和平板浏览器直接用手机 App 界面(#1749,#1757):手机和平板浏览器打开电脑的 H5 地址后,进入的就是 cc-haha 手机 App 的同一套界面,只连接打开它的这台电脑;没有推送通知和锁屏活动,不能添加或切换其他电脑。同时修复 iOS 26 Safari 键盘弹出后输入框被顶出屏幕、拼音输入变成乱码的问题。
  • 手机 App 0.1.3(#1785,#1764,#1770,#1795,#1808):可以填写明确的 服务器地址(不加密,仅限可信网络);Android 扫码失败时可以选相册里的二维码照片配对;AskUserQuestion 的选项说明可以展开;超长消息和重复进入首页不再导致 Android 内存溢出或崩溃;本地存储临时读取失败后,已配对的电脑会自动恢复,不用重新配对。
  • Log the D1 failure class and retry an overloaded read once
  • Defer Codex MCP tools except Computer Use
  • Keep a sent image in place when the host stores the message
Trueforge
  • 0359626: Add a Memory toggle to the agent builder, shown once the tenant has a memory provider. In OSS, admins configure Zep in Settings → Memory (/api/v1/settings/memory-providers).
  • Updated dependencies [0359626]
  • @truefoundry/trueforge-assistant-ui-runtime@0.184.1
  • #1552 d2556bc Thanks @okwasniewski! - Breaking: inputValue() on a node with no value, such as a heading or a button, and isChecked() on a node with no checked state throw ACTION_FAILED, as in…
  • #1524 cba5e0e Thanks @okwasniewski! - The built-in agent can resize the viewport with the new resize_viewport tool, so agent.act and e2e explore can check layouts at desktop, tablet, and phone…
  • #1479 d15a98e Thanks @okwasniewski! - app.baseUrl no longer ends in a slash, so ${app.baseUrl}/api/users requests /api/users instead of //api/users.
Opik
  • [issue-3411] [FE] feat: open any LLM span in the playground
  • chore(deps): Update python:3.12.16-slim Docker digest to e695ab1
  • [issue-9381] [SDK] fix: treat a non-object LLMJudge verdict as a parse failure
  • No notable changes
BrowserSkill
  • MCP server: bsk mcp exposes BrowserSkill over stdio through the existing daemon and extension. bsk install-skill --mcp configures Codex, Claude Code, or Cursor.
  • Controller identity: session startup accepts an agent name, project, and color, shown in the extension's shared-control overlay and session metadata.
  • Date and time inputs: fill supports date, time, datetime-local, month, and week, with format and min/max validation.
  • feat(pages): showcase supported model providers
  • feat(lang): review Starlark files
  • feat(rules): add C# review rules
AI
  • #1973 374fb69 - The adapters now call the provider HTTP API with fetch and do not create the @mistralai/mistralai client. At runtime they load only the SDK error classes. Bundles are smaller.
  • Updated dependencies \[48cbc66, 374fb69]
  • @tanstack/ai@1.2.0
  • Versioned research artifacts
  • list_backtest_runs finds the backtests earlier turns and sessions
  • A research goal's objective can be edited in place (#2093, thanks
  • A body that stands on your desktop. She sits in one window, falls with gravity, hops when you pat
  • Five characters, each with its own voice. Andee itself (the plush ball of the brand illustrations),
  • The same rules on both bodies, because she runs on the phone too — in the ball's own window, so
  • ⚡ Lower terminal memory use: Stop repeatedly reading and parsing font files for terminal ligatures, fixing a source of high memory use and launch crashes in sessions with many terminals.
  • ⚡ Faster worktree switching: Performance improvements make switching between worktrees faster.
  • 📝 See what agents changed: Chat turns show changed files and line counts, with patches available inline on desktop and phone.
  • 安装包和自动更新源迁移到 installer.firemux.site。
  • 更新检查优先使用国内 CDN,保留 GitHub Releases 备用更新源。
  • 同步内置服务的版本检查地址。
  • .Net: Remove migrated .NET Milvus connector
  • .Net: Bump Aspire.Azure.Search.Documents from 13.4.6 to 13.5.4
  • .Net: Bump Aspire.Hosting.Azure.CognitiveServices from 13.4.6 to 13.5.4
  • #4687 2af56d2 Thanks @sailist! - Fix repeated HTTP 400 errors in sessions caused by empty system messages for dynamic tool schemas.
  • #4705 2391a77 Thanks @kimi-agent-bot! - web: Fix session titles not being generated automatically after the first turn of a new session.
  • #4705 2391a77 Thanks @kimi-agent-bot! - web: Fix local images not displaying when their paths contain spaces.
  • Safer Mac settings: failed config saves no longer overwrite secrets, and turning off Tailscale exposure keeps Gateway authentication intact. Fixes #187600, #187597 Thanks @anmol0b.
  • More reliable channel replies: Telegram, Discord, and webhook conversations keep reaching the agent after request handlers or earlier turns finish.
  • Conversations survive more interruptions: oversized turns retain useful history, and stopping the first Claude CLI turn no longer starts the next message from scratch.
Paseo
  • Added a first-run setup on Paseo Desktop and web that walks new users through theme, providers, projects, agent orchestration and the mobile app, with Settings → General → Show setup again to…
  • Added project suggestions from the folders where you ran Claude Code, Codex, OpenCode or Pi, shown in the setup Projects step
  • Changed the theme picker to a grid of tiles with a preview of each theme
  • ✨ Transitions are clips. A transition sits on the timeline and is placed, moved and trimmed like any other clip, over a cut or on a single shot.
  • ✨ Transparent projects. A clear background with a checker under the preview, or a background colour of your choice, and export with alpha as ProRes 4444 or VP9
  • ✨ Chroma key that finds the screen itself, an eyedropper, Remove Black, and batch edits in the inspector
  • #1081 686fee5 Thanks @rstuhlmuller! - Allow MCP hosts to defer unrelated stale Claim pages during focused updates while preserving their evidence debt.
  • #1069 7bff3c6 Thanks @p4norman! - openwiki visualize can open on a specific page: a #page=<id> URL fragment (for example selects that page on first load.
  • 10213067 job postings. See jobs-latest for the current database. Sources and licences
  • Changed agy acp to offer every model available to your account to every ACP client, where previously only a fixed list of editors were offered models other than Gemini.
  • Changed the tabs in /config during --remote-agent sessions from Local and Remote to TUI and Remote Agent, so it is clear that one configures the terminal UI and the other the remote agent.
  • Improved how quickly the agent moves on after a terminal command exits while a background process it started still holds the command's output open, which now takes at most about 1 second instead of 5.
  • Mika-v0.1.37-arm64.apk:近几年的安卓手机基本都选这个,更小。
  • Mika-v0.1.37.apk:通用包,所有手机都能装。不确定就选它。
  • 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
  • 更新前请注意:这一版换了运行时,内置引擎升到 DSH 0.2.2-rc.1。 升级后 Linux 环境里只保留 /root,用 apt 装的软件和放在 /usr/local、/opt 等目录的文件会被清掉,需要重新装。对话和设置不受影响。也因为换了运行时,这次还是下整包。新引擎写入后,定时任务不保证还能被旧版引擎读回,不建议退回旧版。
  • 这一版的引擎升级还没在真机上验证过。 目前在 x86_64 模拟器上用测试模型走过从 1.0.10 升级、发消息、定时任务、窄屏网页版和手机工具,没用过真实模型账号。遇到问题请开 issue 反馈。
  • 网页版:手机宽度下改用官方的抽屉和设置页,顶部的菜单按钮打开官方抽屉。平板和横屏沿用原来的布局。
  • fix(engine): a live component that calls operator.update() before any update_full() no longer loses what it mounted on its first run.
  • fix(engine): a memoized function whose tracked state changed while its memo stayed valid had its target states and nested memo entries deleted at commit; they are now kept.
  • fix(litellm): LiteLLMEmbedder accepts encoding_format="base64" responses (a much smaller payload over the wire), probes the right dimension for them, and refuses dtype options…
Open Connector
  • feat(oomol-rag): add atomic block-level article editing
  • 🔀 The Pull Requests UI is now a compact list with checks and merge actions, and a session's linked PRs stay visible while its host is offline.
  • 🧰 Chats show the actions that an agent took more clearly: tool calls read as plain-language actions with a one-line result and expand to labeled fields instead of raw JSON, and edits and new files…
  • 🖥️ Run sessions on any machine over SSH, launched straight from the desktop app. Local sessions can also open localhost apps in the desktop browser pane, and the Files panel opens files in VS Code…
  • Python steps trust what kane-cli trusts — the certificate bundle is handed to Python programs as SSL_CERT_FILE, so runs behind a TLS-inspecting proxy or corporate network no longer fail with…
  • A broken trust file can't poison a run — if the configured trust file is missing or invalid, it is never passed along to Python programs instead of being handed over and failing mid-run.
  • Locator actions say why they failed — instead of a bare "action failed", the error names the reason, so you can tell a missing element from a blocked click without digging through logs.
  • feat(library): How Do You Build a QuickBooks AI Agent for Month End Close?
  • feat(library): How Do You Build a Notion AI Agent for Pages and Databases?
  • feat(library): How Do You Build an Airtable AI Agent for Data Entry and Reporting?
AgentOS
  • api: a usage event names the account its call ran for (d2b0922)
  • api: a panel call may carry accountIdOf; a rejected promise from it is handled; its failure logs a fixed error type (0550167)
  • api: a revoked proxy from accountIdOf cannot make its failure line throw (0ab1229)
Ruflo
  • browser tools: browser_open accepts only a fixed list of plain Chromium switches. No caller value may add an agent-browser option.
  • memory export / import
  • For MCP callers, memory_export and memory_import paths must be inside the project. Symlinks, .git, executables, and servers started from a home or filesystem root are refused.
  • 💻 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):打开下载的 .dmg 镜像,将 toonflow.app 拖入“应用程序”,复制完成后再从“应用程序”打开。
  • 🐧 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,安装后从应用菜单打开。
  • harbor: record model parameters and agent settings in experiment metadata (2b2b9fb)
  • client: validate required name and annotator_kind columns in annotation dataframes (3146086)
  • commit: 5ab30ec856e57e2ae8944657e9e2d69264c672f5
  • source tag: beta2026-11-05_10-46-18
  • regions: cn global
  • docs: browser extensions, terminal link menu, and a wording trim
  • docs: add Migrate from Conductor and Conductor Cloud guides
  • chore: relicense under MIT
  • perf(ego-browser): collapse click actionability checks into one injected call
  • fix(ego-browser): isolate default screenshot paths across runs in the shared Node process
  • docs: sync README benchmarks and comparison table with lite.ego.app, add a Windows download
  • release(langgraph): 1.2.20
  • feat(langgraph): give each interrupt() in a task its own id
  • fix(langgraph): point the multiple-interrupts error at the current docs
AI SDK
  • 5d18351: fix(ai): detect downloaded audio types from bytes when servers return a generic or non-audio content type such as application/octet-stream
  • 8da48db: Respect recordInputs and recordOutputs when delivering telemetry events to integrations and the ai:telemetry channel.
  • fd7e8a7: Preserve the order of text and complete function calls in streamed Google responses, including standalone thought signatures on empty text or reasoning parts.
  • build(deps): bump boto3 from 1.43.118 to 1.43.123 (9b30fc1b4) — dependabot[bot]
  • build(deps): bump ruff from 0.17.2 to 0.17.3 (247e59a8d) — dependabot[bot]
  • build(deps): bump pydantic from 2.14.1 to 2.14.2 (7d06ef434) — dependabot[bot]
  • You can now export a platform-hosted assistant directly from the backups card in the web interface, making it easier to download and migrate your assistant.
  • Exporting an assistant now correctly registers hosted backups as runtime exports, ensuring the download step completes successfully instead of failing to resolve.
  • Importing a migrated assistant bundle no longer fails when the bundle contains credentials — credentials are safely stripped during import so the process can complete cleanly.

Nov 4, 2026

  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
Orca
  • 🎛️ Choose Chat or Terminal from the start: Onboarding now lets you pick Chat UI or Terminal, then choose an agent and install the skills for that setup.
  • 💬 Keep control of conversations: Preserve chat inputs through uncertain delivery and Stop, choose whether sending moves the scroll position, and decide when idle chat agents stop—or keep them running.
  • 🗂️ Keep projects tidy: Adding a project keeps discovered worktrees hidden, while create and clone handle symlinked parent folders consistently.
Loki Mode
  • A test that was already failing before the change, with the same error, no longer blocks delivery: it gets one fix round, and if it still fails the same way the run is delivered as PARTIAL with…
  • A failure is treated as pre-existing only when its errors match the base run exactly: a changed value inside an assertion (for example a different URL path), a new unhandled-error report, a…
  • Deleting or renaming an installed package now counts as a dependency change, so a resulting "Cannot find package" failure is never passed off as pre-existing.
Ziva
  • Godot 4.2+
  • Active Ziva account
Glimmervoid
  • glimmervoid checkout <repo> gives an agent its own checkout of another repository, by configured project name or path, that is removed with the session when it closes.
  • Session Coherence turns itself on where a repository uses Coherence. On an experimental build, Show session Coherence is now on by default, and it applies only to Claude Code and Codex sessions in…
  • npm i -g glimmervoid now installs exactly the dependency versions each release was tested with.
  • The chat widget lets the AI agent in a visitor's browser talk to your agent
  • Server(session_store=..., customer_store=...) accepts an s3://bucket
  • The chat widget (parlant widget init) takes the agent's picture in one
  • One command puts two agents on a task. One builds, one challenges, and the work reaches you already argued over. Signed in to only one vendor? You still get the pair, and the command says so.
  • Stop watching terminals. 2 need you shows up in your status line, and sno squad lays out the whole squad: who needs you, who is working, what is finished.
  • Work outlives the agent doing it. Quota, crash, closed window: the task comes back, and you get the truth about how. Resumed, rebuilt or lost. A full context no longer costs you the task either.
Phoenix
  • Updated dependencies [6a60c71]
  • Updated dependencies [491ae7a]
  • @arizeai/phoenix-client@7.18.1
Rakazo
  • On iPhone and iPad, an image pasted into a chat's message field is added as an attachment, as on web. Text copied with it still goes into the field. This needs the new app build.
  • On iPhone and iPad, Rakazo appears in the share sheet. Share a link, text, photos or files from any app, pick a bot, and it lands in that bot's message field as a draft. This needs the new app build.
  • More keyboard shortcuts on web and desktop: start a group chat, connect apps, open the chat's details panel, jump to the message box and dictate.
  • Attach files or folders from the desktop composer's paperclip menu, or drag and drop them into a session. Attachments insert local path references.
  • Tutors can show snapshots of your own images and annotate them in the conversation.
  • Import notes… is now Extract files…, with clearer guidance separating graph extraction from attaching material for a lesson.
  • An [[llm_fallbacks]] entry that sets api_key_env for openai-oauth, codex or cursor now fails ai-memory serve startup instead of being ignored. Remove the key to keep the native login.
  • Under --http-stateful, a request without a session now gets HTTP 400 instead of 422.
  • ai-memory mcp-bridge --no-session-id now sends a tool call that names no scope to the project of the directory the bridge was launched from.
AgentOS
  • anthropic: count cache writes from the split when a usage gives it alone; trim a base URL given to the provider directly (967488b)
  • anthropic: set the provider's base URL once in its configuration (35b9a1a)
  • pricing: bound a cached request at a default-only write rate; strip trailing slashes from the Anthropic base URL (5105e12)
  • #818 778fec5 Thanks @alebgl77! - feat: add Mistral as a model provider (MISTRAL_API_KEY, optional MISTRAL_BASE_URL)
  • #1103 959616e Thanks @MohammedAlkindi! - Show the update planner the recent commit messages after each page's baseline, as hints about where to look, so it can route a change by its intent and not…
  • #1102 aa2eaec Thanks @MohammedAlkindi! - Cap the changed paths and Claims issues in the openwiki_begin result, with exact totals and truncated flags, so an update with a lot of drift no longer…
  • Changelog for v0.65.0
  • chore(release): bump version to 0.67.0-nightly.20261020.gb041de017
  • Changelog for v0.66.0-preview.0
Strands Agents
  • feat(docs): add migration guide from strands-agents-tools (f078b5c71)
  • feat(openai): place explicit cache points on GPT-5.6+ by default (663987245)
  • feat(site): import versioned external docs (e4a0ee28d)
  • fix(marketing): say "formerly Superset" in the homepage description and link paradev.com
  • fix(i18n): stop React Compiler from showing raw message ids, and catch it in check:i18n
  • fix(api): flush the reconcile-usage cron check-in before responding
  • Mika-v0.1.36-arm64.apk:近几年的安卓手机基本都选这个,更小。
  • Mika-v0.1.36.apk:通用包,所有手机都能装。不确定就选它。
  • 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
Omnara
  • feat: show agent status as GitHub checks on pull requests
  • fix: report web_fetch truncation only when content is cut
  • fix: reject invalid ports in endpoint URLs
  • docs: overhaul Pantheon recommended models, align zh-CN hero, and add $40 dual-subscription preset
  • refactor(background): dissolve board ceremony — watchdog abort, structured v2 registration, gate trimming (DRAFT, rebase planned onto
  • fix(foreground-fallback): send only the internal reminder on replay
  • LLM summaries, compression, consolidation and reflection need AGENTMEMORY_LLM_PROVIDER=anthropic|openai|gemini|openrouter|minimax plus that key.
  • Cloud embeddings need EMBEDDING_PROVIDER=<name>; local stays on-device, unset means BM25 and graph only
  • OPENAI_API_KEY_FOR_LLM is removed
Ruflo
  • embeddings: embeddings_search honours threshold: 0; CLI reads --db-path, --max-size, --ef-construction, --batch-size
  • session: save/restore/export honour the component flags
  • memory: list pages row identities before loading payloads (#4687, closes
HyperFrames
  • Producer: Retry with fewer workers when every page times out loading (04bc4737ad1d3, #7171).
  • Lint: Flag a data-automation or data-fx-chain the render cannot read (cd7b7383cb19c, #7169).
  • Studio Server: Splitting a clip that starts after another clip cuts where it plays (0bbc9b2943ad6, #7172).
Sim
  • feat(video-generator): refresh native provider catalogs and request protocols
  • feat(image-generator): refresh provider model catalogs and validate requests
  • feat(extension): improve page capture, attachments, and browser shortcuts
  • Side chats — ask about any passage without derailing the conversation.
  • Spend visibility — people see what their work costs; admins can set spend caps.
  • Smoother app connections — reconnect or disconnect apps, Zoom, and large tool catalogs.
  • Let a GitHub workflow trigger watch many repositories
  • Add Devin as a built-in provider
  • Add neutral and exact agent colours
  • Changes since langchain==1.4.8
  • release(langchain): 1.4.9 chore(deps): require langchain-core>=1.7.5 for refusal handling feat(core): standardize model refusal content blocks
  • Improved Agent Panel memory usage by loading tool-call diffs and parsing thread message Markdown only when first viewed, instead of when loading a thread. (#65444, #65470; thanks Vanuan)
  • Improved AI commit message generation by using the three most recent repository commit messages as style examples. (#61528; thanks matheuslrd)
  • Improved CPU usage and disk reads while the agent was working, especially with a long thread history.
E2B
  • a8155c4: Add strict to pause() (default false). With strict: true / strict=True a pause is refused with ServiceBusyError / ServiceBusyException while the sandbox's guest agent does not respond…
  • PostgreSQL 可存放技能包和知识库。多实例共用定时任务、频道、头像、插件和已发布专家
  • Docker 与飞牛增加精简版。存储后端增加火山引擎 TOS
  • 可预建 LDAP 账号,聊天列表可隐藏任意专家
  • #1481 e481624 Thanks @Arnavkar! - fix(core): upgrade the Amazon Bedrock and Google Vertex providers to their AI SDK v6 releases
  • For Gemini models, if GOOGLE_VERTEX_API_KEY is set, Vertex now uses express mode and ignores GOOGLE_VERTEX_PROJECT, GOOGLE_VERTEX_LOCATION and GOOGLE_APPLICATION_CREDENTIALS.
  • Bedrock clamps a temperature below 0 to 0 and above 1 to 1, returning a warning in either case.
  • Docs: legacy .doc import rounds out — nested and floating tables, text columns, page numbering and borders, cell shading, picture crops, fonts and bidi text.
  • Sheets: legacy .xls workbooks keep cell formatting (reported
  • Slides: legacy .ppt import, stage 1 with editable text (@tokg-venv); MathType equation previews render correctly (reported
  • pi-btw 0.8.1. While a /btw side question is being answered, messages you submit in the main editor are no longer held until it finishes; the answer streams in the overlay.
  • ROB-1624 Notify Slack when a release workflow fails
  • Add allowed_tools to MCP servers to load only some of a server's tools
  • feat(alert): add webhook destination
  • [NA] [BE] fix: keep ClickHouse query log comments to plain text
  • [OPIK-9052] [BE] fix: price upper-case model names like lower-case ones
  • [NA] [SDK] [DOCS] Update automatically OpenAPI spec and Fern code
AI
  • Updated dependencies \[5f62940, 5f62940, 5f62940, 9ac67f3, 9ac67f3, 4caefd7, 4caefd7, c7a1b4d, d6ead13, 450950d, 07d2e42, 13b1ca9, 09dde65]
  • @tanstack/ai-acp@1.0.1
  • @tanstack/ai-harness@0.2.0
  • Voice calls: call a session's agent from the session or the agent page, with screen share; the chat keeps what was said (flag voice_call).
  • Speech-to-text and text-to-speech through the LLM gateway.
  • Discord sits on the Channels tab like every channel, marks the message the agent works on, and sets up its app with one command.
n8n
  • Make npm install of n8n work on all platforms again (6abcfc7)
  • Local dev binds loopback — make dev / make start no longer listen on
  • Stored credentials are encrypted at rest — a new env-only
  • The multi-process gate is stricter — GATEWAY_WORKERS > 1 (or
Nocobase
  • [File manager] Fixed audio previews continuing to play the previously opened file when switching attachments or reopening the preview for another file
  • [Backup manager] Fixed missing feedback during large backup downloads by using short-lived links and native browser downloads.
  • #4621 1b13934 Thanks @Grapedge! - Allow /model and /effort while the agent is working.
  • #4544 e24e1e1 Thanks @sailist! - Add support for forking running sessions and forking from a selected historical turn.
  • #4520 3ed0e68 Thanks @liruifengv! - Merge skills from every configured directory in multi-directory projects, with the primary directory taking precedence on name conflicts.
Pydantic AI
  • GHSA-m5vj-8mcw-82qx (high): web_fetch_tool and the WebFetch capability's local fallback re-indented a JSON response on the event loop before applying the returned-content limit, so a deeply nested…
  • GHSA-5fj9-5pjw-vpwc (medium): The cloud-metadata blocklist that URL downloads check did not include Tencent Cloud's instance metadata address 169.254.0.23, so someone who can supply a URL the…
  • GHSA-jp94-mmr3-2pvv (medium): With OpenTelemetry instrumentation enabled for embedding models, embedding spans recorded the extra_headers and extra_body embedding settings, so credentials passed…
Andee
  • You can see it working. A turn waiting on the model used to look stuck — thinking makes no tool
  • It can decide not to speak. A [quiet] turn does its work silently and leaves one line you can open.
  • 「消息都帮我处理一下」 finally works. That instruction never reached the notification triage, which
  • 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.
  • @mastra/temporal@0.5.0: Steps now validate inputSchema and requestContextSchema before running (can fail runs that previously “succeeded” with wrong input); set validateInputs: false to opt out.
  • Experimental memory changes: experimental_subconscious and Memory knowledge option are removed (now throw).
  • knowledge.reindex() is removed; use upsertPage/note writes (and reindexLinks() for link indexing) instead.
  • The manifest of the current job database.
  • B4 now uses the shared init / validate / preview / approve / render workflow alongside B1–B3.
  • Also included: the approved eight-language showcase/FAQ refresh and A6 reference-image upload support. B5 and later are not part of this release.
  • Validation: 102 runtime tests, 20 A6 unit tests, 13 repository regression scripts, multilingual README checks, and an isolated technical render with existing accepted media.
  • feat(ai-gateway): raise Pro/Team/Enterprise gateway limits to 10k req/min and 5 GiB/min
  • feat: improve internal error messages
  • feat(skills): trace Codex skill resource loads
  • 6fabc97e728254ad3e009886e32a816c8f4d00f9 release(notes): keep the v1.3.4 known-issue scope and correct the gate claim
  • d151f4109ed0bd917cf976c7191896590a52c242 release: prep v1.3.4
  • Mistral realtime TTS sends a reply's text over one connection per reply. Documentation.
  • Gemini's automatic_continuation controls resumption of responses that reach the per-request token limit. Documentation.
  • MCP's max_sse_event_size configures the maximum accepted event size for tool results. Documentation.
  • Settings are reorganized into focused tabs. General preferences split into dedicated tabs: Appearance and Notifications move under System settings, while General now shows app information…
  • Import references in batches from DOI, arXiv, or PubMed. Add from DOI or link now accepts up to 100 mixed identifiers per import — DOI, arXiv ID, PMID, PMCID, or official article links —…
  • AI study snapshots for literature. The literature detail view gains an AI snapshot card that summarizes a paper into five structured sections — research question, study population, methods, main…
  • Match the desktop's lattice loader on the live chat turn
  • Add setup-token option to local Claude Code login
  • Detect cached Gemini credentials
  • New enable_file_autocomplete setting (and VIBE_ENABLE_FILE_AUTOCOMPLETE) turns off @ file completion and its background file index.
  • A Jump to bottom button appears over a scrolled-up conversation and returns to the latest message.
  • Ctrl+C closes the PR/MR list, and Home, End, PageUp and PageDown move its highlight.
  • Follow the hierarchy. Expand called Workflow Tools to see their internal execution steps under the invoking Tool span.
  • Understand each invocation. Repeated calls and failures retain their place in the trace.
  • Keep the LLM context. Inspect formatted prompts and responses, token usage, and project, trace, and observation links.
  • Added acp to agy --help: agy acp runs the CLI as an Agent Client Protocol (ACP) server over stdio, so editors can use it as their agent, with images and embedded file context in prompts, steering…
  • Added --remote-agent to agy --help, which runs the session's conversations on remote agents instead of this machine, where --project names a remote project by ID or name and otherwise…
  • Added local projects to /projects, so in a local session you can list your projects with the current one marked, create one from a name and a set of folders given as absolute, ~, or…
  • Improved reliability of the release and build pipeline, ensuring smoother and more consistent delivery of updates to users
  • Infrastructure updates to keep the development and deployment process up to date with the latest tooling
  • Under-the-hood improvements to support continued stability and performance of the assistant
LiteLLM
  • Lens: If your Helm values set lensWorker.enabled, remove it before upgrading so the chart renders
  • Embeddings, images and audio: Added /v1/videos support for custom providers registered in custom_provider_map
  • Lens: Moved the Lens beta out of the proxy package and Helm charts
  • Adds an operator-scoped model.catalogRefresh setting and lets privacy.usageStatisticsEnabled: false stop background catalog refreshes while local metadata remains available.
  • feat(managed-agent): support full model configuration at session creation
  • When experimental.unifiedAgentLoop is enabled, ACP Goal turns run on the shared agent loop with shared rounds, tool batches, and stop points.
  • commit: ac785430449042f120a296b881a25ae52333f1e1
  • source tag: beta2026-11-04_11-03-43
  • regions: cn global
  • No model-provider catalog or request-admission behavior changes are included
  • Add owner-bound, read-only background-terminal status with bounded terminal
Langflow
  • fix(deps): backport filelock 4 upgrade to 1.13.1
  • fix: confine Git loader files to repository
  • fix(oracle): confine loader local file access
Noetic
  • feat: one way to attach MCP servers, served to every agent (2195f69), closes
  • fix: repair CI for main - Updated .github/dependabot.yml:13 to exclude (98839023a) — Vaibhav Upreti
  • Updated bundled Claude CLI to version 2.1.318
  • Ollama support: an ollama LLM profile format for a local or LAN Ollama server. No API key is needed, OPENAI_API_KEY is never inherited, the base URL defaults to and is normalized to the /v1 root…
  • Setup token for first-run setup: while no admin password is set, the server generates a random 128-bit one-time token at startup and prints it to its log…
  • ARTEX_BIND_ADDR sets the address docker-compose.yml publishes the web UI on; the default is 127.0.0.1.
  • Added a record of your answer when Claude Code asks to turn on the mods it writes in a session
  • Added assume_role on mantle upstreams of the Claude apps gateway, for a role in another AWS account and for per-developer AWS cost attribution
  • Added an effort option to a mod's $.model.fork, so a quick side job can run at low whatever effort the session uses; the result reports the effort sent

Nov 3, 2026

  • Gas City: failed bd commands in a rig can be misreported. gc exports GT_ROOT into agent sessions itself, so formulas still resolve.
  • When you run bd by hand in a rig without GT_ROOT exported, the city's formulas ($GT_ROOT/.beads/formulas) and town molecules are not found.
  • 10eec1f707feea8e9472ab0c0ed4ec4ed9dcb05b release(notes): align the v1.3.3 known-issue and gate notes with rc.1
  • Added targeted skill references with skills get <name> --ref <ref> and the MCP agent_browser_skills_get tool's refs argument, so agents can load one reference or template at a time
  • Added Chrome security warnings when newer stable releases list CVE fixes, without blocking commands. Set AGENT_BROWSER_NO_VERSION_CHECK=1 to skip the check
  • Fixed session restarts when commands omit or change timeout settings, preserving open pages and reusing the running daemon's default timeout for waits
  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
Kortix
  • perf(api): stop prompt-queue polling and cron prompt stacking, route S3 around the NAT
  • fix(ci): the dev frontend deploy no longer ships the previous run's build
  • docs(learnings): a private repo's CI never fails over to billed GitHub-hosted runners
Noetic
  • ci: publish prereleases from next to the next dist-tag (0b247bc), closes
Codex Security
  • GitHub Action for Codex Security 0.8.3.
  • Installs @openai/codex-security@0.8.3.
  • yaml uses: openai/codex-security@90c85a3617e06eb5b4ac139778ef06581f44e532 # action-v0.8.3
  • go: same-file member calls resolve through the receiver's type; cross-package pkg.T receivers bind only via the file's import; interface method calls dispatch to implicit implementers.
  • rust: module-qualified calls (crate::b::helper()) resolve cross-file; calls inside expression-macro arguments are linked.
  • kotlin: overloaded and init {}-block calls resolve; unreachable external-inherited calls are dropped.
Glimmervoid
  • Team review works without a GitHub team. If you are not on a GitHub team, Reviews setup now offers Only requests to me: pick one of your GitHub organizations, or type one, and team review drafts a…
  • Boxes on a board can hold a picture. A session runs glimmervoid whiteboard attach <image> (or the whiteboard tool's attach action) with a PNG, JPEG, WebP or GIF file, then puts a box with the…
  • Boxes on a board can be wider or narrower. Drag the small handle at a box's bottom-right corner to make it wider or narrower, and undo puts it back.
  • Microsoft Store app: updates through the Store.
  • Desktop app installed from the installer (.exe / .dmg): no change in this release; it updates with the next bundled release.
  • CLI: hermes update, or re-run the installer one-liner.
AgentOS
  • providers: the retry wait ends when the caller's signal aborts (49ff0ed)
Traycer
  • fix(gui-app): a reading position saved inside a turn survives the steer that splits it (6fc31b4)
  • fix(gui-app,desktop): mod+w in a sandboxed page closes its tile, not the task tab (757d0eb)
  • fix(gui-app): a steer-split restore saves its resolved slice at mount, not only after it validates (81807ff)
  • Mika-v0.1.35-arm64.apk:近几年的安卓手机基本都选这个,更小。
  • Mika-v0.1.35.apk:通用包,所有手机都能装。不确定就选它。
  • 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
CopilotKit
  • This patch release improves A2UI catalog serialization so that composed custom component trees are correctly recognized by agent SDKs.
OpenMausBot
  • Parallel threads: on a MausBot Cloud, memory decides how many run at once, and a turn past that says "Waiting for memory" and starts by itself.
  • Who said what: every line a bot reads now says who sent it and whether it's a request or just information, so a relayed "yes" can't answer a question meant for someone else.
  • Teams: "Looked after by", two simple sharing choices, one inbox list, and refusals that name who can share a bot. A team export now leaves out members' private chats.
AI Memory
No release notes published.
E2B
  • 08f05bb: Fix TypeError: fetch failed / NGHTTP2_ENHANCE_YOUR_CALM on concurrent files.write() calls.
  • 1ad3f76: Template builds hash copy() files in 1 MB chunks instead of reading each file fully into memory, cutting peak memory for large files.
  • Breaking: OpenShell now emits native OCSF 1.8.0 records only, so deployments that relied on 1.1 or 1.3 records now receive 1.8.0.
  • The Podman-only [openshell.drivers.podman].additional_ca_bundle setting from v0.1.5 still works, but it is deprecated in favor of the new [openshell.supervisor.network].additional_ca_bundle.
  • On the MicroVM driver, sandboxes from registry images and rootfs tars now run as the image's sandbox account, as documented, instead of always running as 1000:1000.
  • Add a server audit log to Server settings
  • Show all agent settings for an agent on a joined server
  • build(nix): install OpenBot 0.59.0
  • When you turn on Let bots use this computer in the desktop app, the computer stays connected only as long as the sign-in that connected it.
  • The desktop app's native speech recognition and Local Computer Access now need an HTTPS server, unless the server runs on the same machine.
  • Turning off computer access in the desktop app now stays off even when the app can't save the choice to disk. This covers both remote computer access and the local computer.
  • Opt-in preview config: after CI is green, SHIP finds the PR's preview…
  • @on-par/factory-core/testing exports createFakeGitHub and githubError: a real Octokit answered by an in-memory route table, so tests can fake GitHub with strict types.
  • factory classifier report (and --json as tier) shows trust-tier promotion evidence: configured tier, coverage ratchet, per-package mutation vs floor, human-intervention rate with a caveat…
  • 更新前请注意:这一版换了运行时。 升级后 Linux 环境里只保留 /root,用 apt 装的软件和放在 /usr/local、/opt 等目录的文件会被清掉,需要重新装。对话和设置不受影响。也因为换了运行时,这次还是下整包。
  • 定时任务:新建任务、修改任务的名称、指令或时间,都要单独确认。确认卡会展示完整指令、规则、时区和实际触发时间;完全权限、自动审查和「本对话都允许」都不能代替这次确认。拒绝、两分钟内没确认或没有可用的确认界面时,这次改动不会保存。限制:更新前已经存在的任务不会补做授权,请检查并删除不认识的任务;远程主机要装同样的引擎修复才有这道检查;授权不能代替凭据和工具执行的系统隔离。这项还没在真机上验证过…
  • 动效:弹窗和返回手势跟手,松手后顺着手指的速度由弹簧收住;拉过头有阻尼,回弹途中再抓住会从当前位置接着拖。
Sim
  • improvement(chat): let Chat review organization access requests
  • improvement(billing): open a Chat plan popup on send instead of a static card above the composer
  • improvement(workspaces): move workspace updates behind an application operation and let Chat rename workspaces
  • feat(home): remember and clear the cloud repository selection
  • feat(automations): store picked Bitbucket Data Center repositories as clone URLs
  • feat(automations): list each event source's own event types
  • Update litellm model catalog for v1.15.35
  • Add snapshot and changelog for v1.15.34
  • chat: show a long turn's plan up front (a1f81d2)
  • chat: show a long turn's steps with Stop and I'll finish it (5168b28)
  • chat: show what a long turn is doing and has found, not a step list (aa62c5e)
HyperFrames
  • Engine: A script-load failure names where the script came from and why it failed (dec465949, #7088).
  • Skills: Style templates on the HyperFrames plan (2daedcdaf, #6842).
  • CLI: Diff --offset compares the original at t with yours at t + offset (3f42dc1c7, #7057).
  • [OPIK-7791] [QA] e2e specs for release 2.2.118, 2.2.119, 2.2.120, 2.2.121, 2.2.122, 2.2.125, 2.2.128, 2.2.130, 2.2.131, 2.2.132, 2.2.133, 2.2.134, 2.2.136, 2.2.137
  • [OPIK-8651] [FE] feat: search within a thread in the thread side panel
  • [OPIK-8748] [SDK] fix: LLM judges on gpt-6 and Claude 4.7+/5.x stop sending refused sampling params
Langfuse
  • feat(blob-export): add opt-in trace field groups for the legacy traces export (v3 backport of
  • Parallel researcher subagents in /deepresearch, /lit and /summarize no longer fail with args.tasks[0] contains unsupported fields: output.
  • Pi 1.5.0.
  • Security: Pi now ships undici 8.12.3 and minimatch 11.0.3, so the undici advisories on Pi 1.4.0's pinned 8.10.2 copy and the minimatch advisory (GHSA-g3hv-75wg-m249) on its 10.2.6 copy are gone.
  • feat(allowlist): add HAML, Slim and ERB template support
  • feat(allowlist): review .jsonc files
  • feat(allowlist): add EJS, Liquid and Nunjucks template support
  • A group's admin can set a goal with you, split it into checked stages for its members and run it to the end.
  • The desktop app opens a start window while it starts up: a ring of three arcs around the Main Bot's whale shows real startup progress, first-run setup shows where your data lives and lets you pick…
  • The memory panel lists each memory's change history and lets you restore a single change from the journal.
  • Pi 1.5 support: agents/harness/pi, agents/models/pi-ai and the Pi tools now run on Pi 1.5.
  • Far fewer storage writes while streaming: Pi stores streamed answers and tool output in batches, at the latest every second, instead of on every update.
  • Faster with many sessions: Pi's scheduler no longer rechecks every conversation on each commit, so on SQLite one Pi harness with 2,048 sessions went from 88 to 459 agent runs per second.
  • npm: bash npm install -g base44@latest
  • Homebrew: bash brew install base44/tap/base44
  • Version: 0.3.1 NPM Tag: latest
Trueforge
  • b51ed35: Show a back control before the agent name after Try so the header can return to the agents page.
  • 218c95e: Show a Configure model button that opens Settings when the model catalog is empty, instead of a Select model dropdown with a ? monogram.
  • e3fed89: Let users without manage access open an environment in a read-only drawer to view its configuration.
  • docs: changelog and blog post for v1.1.0
  • docs(blog): ChatGPT sign-in recording in the v1.1.0 post
  • fix(parsers): strip UTF-8 byte-order mark in markitdown text files
Esp32 Iphone Agent
  • esp32agent CLI 0.1.0. Install: see cli/README.md.
  • b77b0694318118dcda816667fb974cc65fdcfbb3 MUL-5491: feat(metrics): add MiniMax-M3 standard-tier pricing
  • 98eb195b046ec16290016c31305f1e8878b81981 MUL-6991: fix(skills): advertise filesystem-allocated skill names
  • 2115a513a375bfe2910b76725b617504faee4b25 MUL-7686: feat(runtime): add native ZCode protocol support
  • commit: a5bd13c69b9f984ebee5fe32c737e489b6ab47dc
  • source tag: beta2026-11-03_17-49-58
  • regions: cn global
  • Haystack now requires anyio>=4.14.2. anyio is installed transitively through httpx and openai; versions before 4.14.2 are affected by CVE-2026-63374 (GHSA-82r6-8w77-94w6).
  • Haystack now requires certifi>=2026.10.19. certifi is installed transitively through httpx and posthog.
  • Native startup: Removes Node.js runtime initialization and JavaScript module loading from standalone CLI invocations.
  • More efficient computation: Uses native code, explicit numeric types, and contiguous arrays in text encoding and vector computation.
  • Fewer data conversions: Reduces intermediate copies and temporary allocations.
  • Nothing to do. The browser extension updates itself. A chat that is still answering delays that by at most 30 minutes; a tool call or a message being sent still finishes first.
  • Move chats between projects. Drag a chat onto a project, or onto one of its groups, or use Move to project in the chat's menu. A notice says where it went and offers Undo.
  • The new project's folder, said once. After you move a chat into a project, your next message there tells ChatGPT the project's folder once, so it works in the right place.
  • Insert known agent commands from replies into the Composer draft
  • Reuse running OpenCode v2 services
  • Kept terminal programs running across daemon restarts, crashes and supported updates
  • fix: validate request headers and randomize CDP identities #2342
  • fix(fetch): enforce CORP before redirect and response delivery #2286
  • fix(stylesheets): validate integrity before installing CSSOM #2264
  • Fleet-wide lane limits: GatewayRungDispatchPolicy.fleet…
  • 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.
  • The ball goes into the edge instead of parking on your screen. Left alone for 10 seconds, it slides
  • Everything still works while it is docked: tap talks, double-tap types, long-press opens the card.
  • A fresh install now opens on the conversation, not on an empty stage board.
  • chore: relicense under a modified Apache 2.0 license
  • feat(triggers): persist provider sources and version event delivery
  • fix(docker): include trigger contracts in build stage
  • 🔄 Continue after usage limits reset: An optional setting resumes your chat automatically when its usage limit resets.
  • 📱 Richer reviews on your phone: Browse branch commits, jump between files in diffs, and read or add pull-request review comments alongside the code.
  • 🎨 Make Orca yours: Discover more themes, preview fonts in their own typefaces, and import Ghostty or Kitty configuration files.
  • Release notes — formatted for people.
  • Changelog — plain Markdown for AI agents and tools.
  • npm package
Loki Mode
  • A Loki PR title now names the work, for example "[PARTIAL] Retry utility and recovery handlers (#21 s2)", instead of "[PARTIAL] Loki 10: PARTIAL (e10-<run id>)".
  • On a split issue (loki start owner/repo#N or loki issues run), a slice's working copy is no longer deleted while Loki's background deep verify is still running in it, which left the…
  • A stacked slice PR says when the repository's pull-request workflows skip its base branch, and that GitHub retargets it to the default branch once the PR it is stacked on merges.
  • on_failure option for HookMatcher: Control what happens when a hook callback fails.
  • Updated bundled Claude CLI to version 2.1.317
  • Tests: e2e tests now accept either Bash or PowerShell tool in shell-dependent tests, fixing intermittent failures on Windows
  • Opt-in archive and restore for named plans. set-active-plan.sh --archive <id> (PowerShell: -Archive <id>) previews the move to .planning/.archive/<id>/, --yes (-Yes) applies it, and --restore <id>…
  • When a project holds several named plans and no PLAN_ID is set, hooks inject nothing. The multiple-plans notice now names the archive command.
  • PowerShell inventory modified_at values no longer shift by the local UTC offset.
  • Added a warning to /schedule when a new cloud scheduled task gets connectors you did not ask for, with how to remove them
  • Added a check that asks before a web or shell tool call sends your account email address to the network
  • Added onFailure: "block" for Agent SDK hook callbacks, so a callback that fails blocks the action instead of letting it run
  • Protect the gh-aw canonical session capture from agent tools
  • Add multimodal notes to AskUserAnswer
  • Count BrowserUse model calls toward the host run's usage and UsageLimits
AI SDK
  • 857ca45: Restore server-defined tool metadata, dynamic status, and title when validating tool approval continuations so approval policies receive the same tool definition fields as new model calls.
  • 4d6d830: Support a single telemetry operation for durable WorkflowAgent runs, including the prompt, final output, steps, and tools.
  • Updated dependencies [fa15312]
Omnara
  • feat: add list_github_repositories tool and Git credential guidance
  • fix: preserve why a machine was deleted
  • fix: state run_command's 30s wait cap and storage_upload's exact /artifacts path
  • feat(account): sign in and connect GitHub over SSH and on servers (30be7de67) — Vaibhav Upreti
  • fix(install): fetch the main build again when a republish splits its archive and checksum (783c431e7) — Vaibhav Upreti
  • feat(analytics): record the Connect GitHub gate's exposure, rechecks and exits (faedad0a9) — Vaibhav Upreti

Nov 2, 2026

  • macOS (-sea assets): unsigned (Apple secrets not configured)
  • Linux (-sea assets): SHA256SUMS-sea published with the release
AgentOS
  • voice: StreamingSTTChain reports backups that fail in a switch, and every failure when none is left (ade8590)
Deep Agents
  • Mark interrupt cancellation messages with lc_source.
  • Clarify workspace conflicts, MCP login failures, and disabled Slack MCP access during login.
  • Clarify how /reload refreshes MCP configuration.
Loki Mode
  • haiku and --cost lean now turn on the router's auto preset (LOKI_ROUTER=auto): Haiku runs implement and fix, and a failure the code owns climbs to Sonnet, then Opus; it never steps back down.
  • On bench v2 (24 tasks x 2 reps): the new --haiku passed 48/48 at $0.0433 per passed task, against 47/48 at $0.0574 for the old --haiku…
  • LOKI_MODEL_DEVELOPMENT and --model now override the preset like LOKI_CLAUDE_MODEL_DEVELOPMENT.
HyperFrames
  • CLI: hyperframes audio --read describes audio stems through your HeyGen account, with no Gemini key needed (acd5c30e60, #6993).
  • CLI: reverse start records a reverse run, and cloud render carries its session (7d3c225207, #7016).
  • CLI: reverse start stops a Free account when the server blocks it (7da583ec69, #7017).
AI SDK
  • 51ca5b7: fix(ai): support cross-realm Uint8Array and ArrayBuffer in DefaultGeneratedFile
  • 51ca5b7: fix(ai): handle cross-realm Uint8Array and ArrayBuffer in data-content and file-part-data
  • 51ca5b7: fix(ai): support cross-realm binary buffers and circular structures in cloneModelMessages
  • Deprecated createFileOps, computeFileLists, formatFileOperations, upsertFileOperations, isUrlSchemePath and FileOperations in @oh-my-pi/pi-agent-core/compaction.
  • Context estimates for a model without a known tokenizer family (Gemini, Grok and the like), and token counts made without a model (compress metrics, the legacy estimateTokens export), count text…
  • Context estimates no longer charge a reader another model's thinking signatures or redacted thinking, which the request strips: a Claude session switched to Grok or Gemini, including Gemini…
  • A member call x.m() binds by the receiver's type, now typed from function parameters and class fields and not only locals (void render(Base* b){ b->draw(); } resolves to Base::draw).
  • An out-of-line or split-header definition attaches to the unique owner whose class file is in its include set, instead of cross-binding to an unrelated same-named function.
  • A local callable value (a lambda or std::function) is kept out of global call resolution by byte-range binding lifetimes, so it no longer mints a spurious INFERRED call to a same-named global.
  • Mika-v0.1.34-arm64.apk(约 80 MB):近几年的安卓手机基本都选这个,更小。
  • Mika-v0.1.34.apk(约 92 MB):通用包,所有手机都能装。不确定就选它。
  • 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
Qwen Code
  • Managed memory availability now respects the memory.enableManagedAutoMemory setting.
  • Size-triggered microcompaction now clears old tool results toward a low watermark, so it no longer rewrites another part of the conversation prefix on every subsequent turn.
  • chore: bump version to 0.8.2
E2B
  • 8cc9ed1: Document that logs.stdout, logs.stderr and OutputMessage.line hold raw chunks with newlines kept (join them with '' to rebuild the output), and that Result.png, Result.jpeg and Result.pdf…
  • 7d10433: Fix readLines quadratic line reassembly when streaming long single-line outputs (e.g.
  • ac0727b: Fix OutputMessage.timestamp in the JS Code Interpreter to use the backend timestamp (Unix epoch in nanoseconds) sent with each stdout/stderr message, instead of the client clock…
BeeAI Framework
  • tools: GoogleSearchTool has been removed as Google's Custom Search JSON API is
  • adapters: add MiniMax regional (CN) endpoint selection (f1e1944)
  • adapters: default MiniMax text models to M3.1 Flash (c31de98)
  • build(nix): install OpenBot 0.58.0
  • Keep the modes of a hosted release after a boat wake
  • Show the Usage chart again
  • Merge pull request #268 from palmier-io/fix/email-signup-provision
  • [fix] Omit missing profile fields when provisioning an account
  • Merge pull request #267 from palmier-io/cursor/email-sign-in-c936
Phoenix
  • Updated dependencies [ead7dd8]
  • @arizeai/phoenix-client@7.18.0
Claude Mem
  • Turning off error reporting now also stops errors already saved on disk from being sent.
  • More secrets are redacted before they are stored: Slack and Discord webhook URLs, Stripe webhook secrets, PyPI, SendGrid, DigitalOcean, Shopify, Docker Hub and GitLab tokens, Basic credentials…
  • Other websites running on localhost can no longer send writes to the worker through the work-state, dedup, corpus and advisor routes.
Solidcraft
  • The browser version shows a design again instead of a black screen. A shader that the desktop app accepted was refused by browsers, so the 3D view drew nothing.
  • Agent chat asks more often. Agent chat answers an agent's permission question by itself only when the whole title is an exact tool name of the chat's own server, named the way the Claude adapter…
  • With the Claude adapter, nothing changes.
  • Microsoft Store app: updates through the Store.
  • Desktop app installed from the installer (.exe / .dmg): no change in this release; it updates with the next bundled release.
  • CLI: hermes update, or re-run the installer one-liner.
  • Import an existing recording: use the smaller folder button beside Record to select a local MP4 or WebM, such as a downloaded Teams demonstration.
  • Check actual video and audio decoding before proceeding. Unsupported codecs and invalid media are rejected with an explanation.
  • Add optional details to explain the task, supplement the demonstration, or suggest where it appears. Leaving details empty is supported. Timestamps guide the AI; they do not trim the recording.
  • Files storage limits per project and per account, by plan: Free 5 GB, Credits 50 GB per project and 250 GB per account, Enterprise 100 GB per project and 1 TB per account.
  • Files storage is billed in credits at cost, like compute: project Files, each session's workspace and scratch disks. Billing shows Storage per project next to compute.
  • The Files page shows a project's storage against its limit and warns before a folder is too large to mount in a session.
  • Docs: legacy .doc/.dot files open as converted .docx copies — text, styles, tables, lists, pictures, text boxes, footnotes, comments, headers/footers and multi-section documents (@tokg-venv, asked
  • Sheets: live values for external-workbook links and a sheet links tool in the CLI and MCP (@tokg-venv); Queries & Connections pane and sideways wheel scrolling on a wide ribbon (@DuBearte, asked
  • Slides: context menus stay usable on small screens and flyouts scroll in place; setText keeps paragraphs together (@tokg-venv).
  • > [!WARNING] > Python 3.10 support will be dropped in langchain-core 1.8.0 and langchain 1.5.0. The current minor versions are the last to support Python 3.10. Please upgrade to Python 3.11 or later.
  • Changes since langchain==1.4.7
  • chore(langchain): bump langchain-core release(langchain): 1.4.8 chore(deps): bump pyasn1 from 0.6.4 to 0.6.5 in /libs/langchain_v1 chore(deps): bump joserfc from 1.7.4 to 1.7.6 in…
  • [PR] (ai: anthropic/claude-opus-4-6) fix(release): nuclear box = deps download + build + publish, zero checks; add local zero-install publish script
  • [PR] Manually remove the fucking AI restrictions | 手动取消傻逼ai限制
  • [PR] Restore the old CI release manual. | 恢复老的ci release manal
  • fix(tool): name missing paths in file_read_diff not-found error
  • fix(session): print an empty items array from session show --json
  • fix(action): keep OCR runtime stable during review
  • Pi joins as a fifth agent framework. Alongside Claude Code, OpenCode, Codex, and CodeBuddy, you can now run sessions with the Pi agent, including onboarding, settings, connection validation, and…
  • Add references straight from a DOI or arXiv link. The literature library gains an Add from DOI or link entry: paste a bare DOI, a DOI resolver link, or an arXiv abstract/PDF URL, review the…
  • Agents annotate PDFs while you read. Reading agents can now list, locate, and create notes and marks in the literature library's existing Notes & Annotations for linked PDFs, so requests like…
  • npm: bash npm install -g base44@latest
  • Homebrew: bash brew install base44/tap/base44
  • Version: 0.3.0 NPM Tag: latest
  • One agent can ping another while you are typing in that window. Your half-written message stays where you left it.
  • Nothing gets sent on your behalf. The ping waits in the agent's inbox until it looks.
  • It works the same in tmux and in Orca.
  • @mastra/playground-ui: Metrics chart card API cleanup (e.g., removed ChartCard description/summary, removed several MetricsCard.* components, changed related table/legend APIs).
  • @mastra/playground-ui: Removed fluid-hover utilities (useFluidHover, FluidHoverHighlight) and CommandList.highlightClassName.
  • @mastra/core: Removed runtime-importable subpaths @mastra/core/network/vNext and @mastra/core/telemetry/otel-vendor (they pointed to files that no longer ship).
  • A first run that is a story, not a checklist (Android). Language and sound, Andee introducing itself, three
  • Andee says it out loud, in its own voice, on ten recorded lines — recorded rather than synthesised
  • Andee's stage (Android). The phone's unfolded card shows a page Andee writes and keeps, where the
  • Extension update changelogs: see what changed in installed extension packages on the next fresh launch after an update.
  • Faster streaming: large tool calls are parsed incrementally, and code-heavy answers and tool rows reuse highlighting and rendered Markdown instead of repeating the work on every update.
  • Custom self-update commands: use PI_SELF_UPDATE_COMMAND or selfUpdateCommand to update pi through a wrapper or another package manager. See Custom update command.
  • [issue-9080] [SDK] fix: log run_simulation traces to the given project_name
  • [OPIK-9041] [SDK] fix: keep user fields named start_time / end_time / last_updated_at unchanged on replay
  • [NA] [DOCS] fix: align the queue env vars with config.yml and correct the pytest plugin typos
  • A Models page. Configuration → Models replaces the settings modal's AI Model tab as the place
  • A queen writes her own skills. write_skill now works when a queen is alone with you, and takes
  • A skill for everyone asks you first. write_skill(scope="global") writes nothing. It saves a
Reasonix
  • [Desktop · CLI] python -v 不再被当作只读的版本探针 — Python 里 -v 是显示模块导入过程,不是版本参数,所以 python3 -v 脚本 会运行这个脚本。Reasonix 过去把它当作只读命令,在「询问」模式和只读模式下都不经询问就直接运行。现在 python、python3 只有 --version 和 -V…
  • [Desktop · CLI] python -v 不再被当作只读的版本探针 — Python 里 -v 是显示模块导入过程,不是版本参数,所以 python3 -v 脚本 会运行这个脚本。Reasonix 过去把它当作只读命令,在「询问」模式和只读模式下都不经询问就直接运行。现在 python、python3 只有 --version 和 -V…
  • Group chats open again in a team imported by 0.2.2 or earlier. Those imports did not copy each group's own conversation, so a click on the group did nothing.
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)"…
  • #1047 0c41be9 Thanks @naveecode! - fix: stream Anthropic (direct and Vertex) requests so OPENWIKI_MAX_OUTPUT_TOKENS can go above ~21,333 tokens without the Anthropic SDK rejecting them as…
  • #1075 e0627a6 Thanks @jrodeiro5! - Stop a repository worker after five plan or page submissions instead of letting it resubmit a rejected payload until the recursion limit.
  • #1065 6b30dea Thanks @jkennedyvz! - Enforce the docs-only write boundary for repository chat, preventing source-file mutations while preserving targeted reads.
  • Android 17:不打扰模式能点了。系统删掉了后台屏下发输入要用的隐藏接口,root、Sui、Shizuku、自连 ADB 四条通道都只能读屏、点不了,现在改用还在的接口。我们手上没有 Android 17 真机,遇到问题请开 issue 反馈。
  • Android 17:自连 ADB 用的 TLS 库(Conscrypt)升到 2.8.0,在 16 KB 内存页的设备上也能加载。
  • 不打扰模式:新增「保持 X 通道」开关,默认关。打开后,通道断开、App 被杀或手机重启,都会自己连回去。root 打开后 DSH Go 会自己再向 root 管理器申请,给不给仍由管理器决定;没打开就绝不会自动执行 su。这个开关还没在真机上验证过,遇到问题请开 issue 反馈。
  • [REALLY COOL!!!] The official phone-harness discord server is now live! I'd love to talk and build with you guys!
  • Send files and links to your iPhone: new push_file and open_url helpers put a file on the iPhone or open a URL on it directly.
  • Smarter setup: agents check what your machine can run and pick a phone that works, falling back to a cloud phone when there's no Mac or local device.
OpenViking
  • Rules(Beta):新增 Rule 服务,并向 Agent 暴露 openviking_rule_search 工具,附中英文指南与 API 参考(#6364、#6613)。
  • 存储与路径锁:新增事务型 SQLite path lock、按 scope 批量刷新 Redis 租约、拒绝跨 scope 的 Tree 锁与路径别名绕过、令牌丢失后撤销租约授权、Provider 迁移后恢复持久化交接;修复递归删除导致 PathLock 提前失效、中文路径旧锁交接恢复;新增可选 Redis 目录 TTL、多写 IOPS…
  • 鉴权与账户:支持用户自助轮换 API Key;管理员列表只显示调用者自己的 user key;校验身份角色;支持角色降级与大小写不敏感输入;Group 成员改为批量添加(#6649、#6610、#6752、#6714、#6796)。
  • 🤝 Orchestrator V3: A lot more token efficient and reliable. Works with both CLI agents and Chat UI, with visual worker indicators in Chat UI to help you follow progress.
  • 🔕 Quieter notifications: Orca now filters out worker agent notifications during orchestration.
  • 📱 A refreshed Orca Mobile: Choose Light, Dark or System and pick your display language in Settings. Enjoy cleaner screens, smoother sheets and faster workspace lists.
  • Window frame: the sidebar back in the app's card, and the rail in the bar's colour
  • fix(redact): mask GitLab personal, deploy and runner tokens
  • Accept Copilot Auto when the session leaves the model currentValue un…
  • 账号支持 TOTP 与 Passkey:二次验证、无密码登录和恢复码
  • 新增 ragflow 知识库连接器
  • 记住登录会自动续期,改密码会退出其他设备。缺 MFA 表的旧库会自动补齐
  • fix: remediate GitHub code scanning finding #2426 - Fixed tests/shared/ (dbfee07d2) — Vincent Hus

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.