AI agent releases: Oct 12 – Oct 18, 2026
545 stable releases from 150 tracked agents and frameworks in the week of Oct 12, 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.
Oct 18, 2026
- Calendar: a new Calendar screen in the sidebar shows your week (an hour grid, all-day events on top, today and the current time marked) or your month.
- Interviews in step with Google Calendar: changing an interview's date in Dossier moves its event in Google Calendar, removing the date deletes the event, and setting one adds it.
- Interview times in your time zone: when a reply said "let's meet at 13:00", the interview could land hours off.
- Turn context. A package registers a factory in the hive.turn_context entry-point group,
- Recall of past conversations reads Chinese. It only took Latin letters and digits as search
- A failed model call stays out of chat apps. A model gateway built on Hive can answer a
- Updates install again: clicking Restart in the update dialog now installs the update.
- Update prompt: an update that finishes downloading while the window is closed is now offered the next time you open the window, instead of being dropped.
- Bug fixes and reliability improvements
- Release notes — formatted for people.
- Changelog — plain Markdown for AI agents and tools.
- npm package
- send: say when Claude Code will hold a message posted into a session
- Support DeepSeek Harness (dsh)
- Follow-ups from the review of #21 and
- Orbit: fixes orbiting about a picked point, which turned the camera instead of the part in 0.16.0 (reverts #2142).
- The failed-tool-call correction sent to DeepSeek models now also covers chat-template tool-call tokens written out as text, not only malformed DSML
- Fixed Ollama chat history estimates counting stripped native thinking and missing foreign thinking replayed as text, which skewed compaction after model switches
- Added nativeJudgmentImageBounds(): the image bounds a native judge enforces, its wire's transport limits tightened by the catalog's model bounds
- feat(analytics): CI-only PostHog key; dev builds never report
- test(server): drop avoidable waits from the slowest server tests
- fix(server): make set_title name an unnamed workspace in the same call
- Automations that react to your apps. Tell a bot "when I get an email from Acme, draft a reply" or "when a new message arrives in #sales, summarize it".
- It sets it up on your connected app, filtered at the source so other mail never wakes it, and shows it in the chat with Pause, Edit and Undo.
- If the app isn't connected yet, the bot asks you to connect it first.
- Nothing to do. The browser extension updates itself once no chat is busy.
- Edit your earlier messages. Click the pencil (Edit message) on any message you sent, change it and send: the chat continues from your edited message, just like editing in ChatGPT.
- Branch in new chat. Answers have a Branch in new chat button (the branch icon next to Copy). It opens a new chat for your next message.
- Unified SDK: vectorless RAG. Local or cloud. Index or chat.
- Faster: trees are derived from layout stats.
- Cheaper than Vector Index: LLMs only write node summaries.
- perf(task-reply): drop the dead v2 answers arg and its caveat sentence
- fix(v2): settle revived children after compaction
- fix(foreground-fallback): walk the chain from the head and skip only tried or abandoned providers
- 0df833ed76c74e25a4c6e12da281cfd0dcb8e17c feat: add basePath to serve Casdoor under a sub-path
- 5619d6c643195fcee14442b7093541d6a8054d46 feat: restyle toasts as plain cards with colored icons instead of tinted backgrounds
- 178a8257ee4e18b264d1fc845da67487790d2295 feat: rotate Marketplace themes on the sign-in page of demo sites
- 两台 Octop 可以连起来用:通过云端协同使用对方的专家,能聊天,也能改对方工作区里的文件
- 专家可以组队。主持人负责分派任务和收尾总结,成员的回复会回到聊天,也会同步到即时通讯
- 新连接器:企查查、OpenAlex、语雀、Obsidian、Agent Mail,以及钉钉命令行(设备码登录后可查通讯录、日程和文档)
- pi-subagents 0.79.1.
- Sessions that have not delegated yet no longer send the bg_wait and subagent_supervisor tool definitions with every request, about 1,500 characters saved per request.
- When a workflow's main process dies, research subagents that had already finished keep their results instead of being marked failed.
- briefing: show a first-week card on Home instead of the getting-started brief (fd53d82)
- chat: group chats by topic on the Chats page (3c6497e)
- chat: keep a long topic list short with areas and quiet bands (88bbb78)
- release v0.1.201
- ci: release PRs default to the full matrix — kill the label/merge race
- image-restore: conversion failures stay fail-open but WARN once per signature (#2824 class visibility)
- The golden image of the inputs ce0e2cde6174, built by the golden image workflow from ef69f5e19c1b46e41e743234fa5019e39442adcb.
- Plugins work from a stock CLI download: install a channel or tool plugin by name, with no Rust toolchain.
- Tool selection: Full stays the default; minimal is an opt-in [tools] setting.
- Authenticated principals reach every entry point: RPC, sessions, tools, and memory resolve to canonical principals.
- The cosign claims were validated
- The signatures were verified against the specified public key
- test(mcp): fire MCP client test deadlines on conditions instead of wall-clock time
- fix(slack): acknowledge a message before the thread lock and session load (134ec1b04) — Vaibhav Upreti
- feat(gateway): use the acting member's own Gmail and other Pipedream accounts (49a6612ec) — Vaibhav Upreti
- Fixed conversation history not showing in Remote Control in daemon mode.
- CLI archives built from the Studio release studio-v2.44.0.
Oct 17, 2026
- Fixed every request failing behind a TLS-inspecting proxy that refuses compressed uploads: a 400/403/415 whose body is not JSON, or the Anthropic API's "request body is not valid JSON" after a…
- Anthropic's request-encoding-rejection now also matches the API's "request body is not valid JSON" 400, which a gzip body gets when a proxy drops its Content-Encoding header
- Removed display() from eval in both Python and JavaScript; print (and console.log in JavaScript) covers it
- Usage ping, removal on request: to have the rows kept for your install removed, send its ID (Settings → About shows it) to dossier.app@pm.me, with DOSSIER_TELEMETRY=off set first so nothing new…
- Demo data: the sample profile, its reference and every example in the README are now Jane Doe and John Doe at example.com, so nothing in Dossier's screenshots, recording or seed data can be…
- #3172 The end-of-run KPI sweep no longer burns the GitHub REST rate limit. It used to make about 800 calls per run.
- kpis: stop the end-of-run KPI sweep from burning the GitHub REST rate limit
- ruflo init / init --dual / @claude-flow/codex init no longer write a permissive .codex/config.toml by default.
- The MCP bridge on loopback no longer accepts tool-call requests from an arbitrary web page: local-only mode now requires the Host to be loopback (blocks DNS rebinding) and the Origin to be…
- The policy ledger now authenticates a retained/pruned ledger against its independent on-disk archive anchor instead of trusting the mutable state.json alone, and streams full audits through fenced…
- Channels. A package installed next to Hive can register a chat-app integration in the
- hive-cred, from the desktop runtime. A script fetches a credential when it uses it
- The credentials tool keeps collect, attach and detach. Its browse, inspect and
- @littleblacklb, for the progress-attestation workflow report and safety requirements that prompted the review. The contribution is recorded in CONTRIBUTORS.md.
- feat(library): 6 Best Relevance AI Alternatives for Flexible AI Agents in 2026
- feat(library): How Much Does It Cost to Build and Run an AI Agent?
- feat(library): 6 Best Make.com Alternatives for AI Agents in 2026
- 独立入口在装有第三方 bundle 的 profile 上启动失败 — #1577 把后端装配放到了延迟根守卫之前。profile 的 dsh.profile.bundles 里有 @deepseek-ai/* 和 dsh-tui 之外的包时,装配会提前解析组合根,根能力守卫立刻装上且撤不掉;DSH 0.2.0-rc.2 的基础行在…
- 更新时不再把其他平台的 Claude SDK 包拉进 profile — pnpm 更新会物化所有平台的 optional dependency;0.15 → 0.16 曾把七个 @anthropic-ai/claude-agent-sdk-* 包(约 1.6GB)装进 profile store,而实际只会链接当前平台的那一个。现在更新和按需安装内核都会按本机…
- 从 v0.16.0 升级:程序内 /update,或
- Automations: set things up by just saying them. Tell a bot "every weekday at 9, send me a summary of my inbox" or "whenever our sales form sends a lead, draft a follow-up".
- Automations in the bottom menu lists them all as plain sentences, each with whether it's working, when it last ran and when it runs next.
- Anything that needs you is at the top, with the button that fixes it. A calendar view shows the week ahead.
- Improved agy plugin validate to also check a plugin's skills, subagents, and rules: for skills, it errors on a SKILL.md with missing or invalid frontmatter and warns when name or description is…
- Improved the inline diff preview when the agent asks to edit or create a file so it renders like the /diff file view, with syntax highlighting for your color scheme, background tints on added and…
- Improved CPU usage while typing and while the agent is working by rendering the prompt area once per frame, reusing measured line widths for unchanged rows across frames, and rendering the…
- Sessions with background work keep working after a restart: a session that had finished its turn but still had a background agent, a monitor or a scheduled wakeup in flight now counts as working…
- A restart keeps a guarded session's scratchpad: a guarded session's temporary folder, which holds Claude Code's scratchpad and background task output, is now named after the session and survives…
- Sessions can clear out their own scratch: a guarded session's temporary folder, where Claude Code keeps its scratchpad, lives under the Glimmervoid folder rather than the system temp folder, so…
- The golden image of the inputs b38d7171e8ec, built by the golden image workflow from ef69f5e19c1b46e41e743234fa5019e39442adcb.
No release notes published.
- Kortix Capture records only the apps and websites you include: an organization policy plus a per-computer list.
- kortix.yaml version 3 agents can opt in to the project repository.
- kortix apps works again from CLI versions before 0.14.0. Since 0.14.0 every Apps command from an older CLI answered "Apps is not enabled for this project".
- ci: add emergency fast mode to carve-out dispatch (skip CI lanes + full test, keep guards + build sanity)
- carve-out publish point for v0.1.199 (tree = master@74bdffd0)
- refactor: lane-decisions reuses the shared RUN_CODE_TOOL_NAME for the PTC omnibus check
- The cosign claims were validated
- The signatures were verified against the specified public key
- fix(responses): backport #46099 and #46109 to stable/1.103.x for v1.103.5
- Cards keep the agent who did the work when new updates arrive, instead of showing Unassigned work.
- Use as mask (⌃⌘M): select two layers, and the one behind becomes the mask of the one in front.
- Masks in the inspector: click + under Masks, pick the layers that should cut the selected one, then press ⌘↵ to apply or Esc to cancel. Each mask is listed with a button to remove it.
- Timeline: masks show up under the layer they cut, marked with a mask icon. Their layers keep their own position, timing and keyframes, so a mask can move, grow or come in late.
- A new icon set: every command has its own icon, in toolbars, menus, the command palette, the browser and the timeline.
- Modelling: sweep along guide rails and two-rail sweep; Class-A surfacing part 3: blend curves, curved-edge fillets, the G2 corner and continuity analysis; Split Body by a plane lying on a face.
- Assemblies and motion: cam and follower joints; motion analysis charts and collisions along a study; export a study or a joint drive as an animated GIF; a clearance report.
- macOS (Apple Silicon): signed and notarized DMG. Drag LearnHarness to Applications.
- Windows (x64): run the installer. It is unsigned; SmartScreen may show an unknown-publisher warning.
- Linux (x64): make the AppImage executable and open it, or install the Debian/Ubuntu .deb.
- #58 8ca326c Thanks @PazChen! - openfilm capture <url> <file.png> takes a picture of a live web page in the browser OpenFilm draws films in, with nothing to install: a product or a site shown as it…
- #58 8ca326c Thanks @PazChen! - An official skill, replicate: give the agent a video and it makes yours like it, keeping its structure, rhythm, look and motion while the content is yours.
- #58 8ca326c Thanks @PazChen! - Saved: your own library, the same in every project. The bookmark button beside Upload switches the media pane to it, and back.
- fix(cdp): let replacement navigation bypass inherited load waits #1509
- fix(cdp): complete Fetch configuration updates during navigation #1582
- refactor(cdp): unify Page access and optional owner lookups #1612
- Show a honeycomb of agent faces on the website
- Add a contributor assignment agreement and a CLA check
- Delete mailbox command receipts by key
- The managed engine's service runs the Reason CLI that set it up
- Reason Network 0.10.5: escape systemd values, skip an empty Reason CLI path
- Pi 1.3.0.
- Tool calls whose arguments contain a backslash sequence like \usepackage no longer lose fields, which mattered for LaTeX in drafts and papers.
- Gateway errors such as "Unable to forward upstream stream" and "service is temporarily unavailable" are retried instead of ending the turn.
- Native chat (experimental): Review turn-by-turn edits inline, expand diagrams to pan and zoom, share chat visuals with a click, and see Claude goals and Codex context usage.
- Shared skills: Shared-skill installation now supports 20 additional agents, with broader WSL discovery and the right install locations for tools such as Kiro, Qoder and CodeBuddy.
- Phone: Workspace filters, grouping and sorting now follow the desktop’s shared choices.
- CLI archives built from the Studio release studio-v2.43.0.
- GPT-6 models (such as gpt-6-sol and gpt-6.1-sol) work on OpenAI and Azure OpenAI-compatible endpoints. Before, every request was rejected because Cline sent max_tokens instead of max_completion_tokens
- The Cline provider's default model is Claude Sonnet 5.5 again, which is back at the top of Cline's recommended list
- Refreshed the model catalog. Adds the CommandCode and Reflection providers. Default models change for 302.AI (Gemini 4 Argon), Eden AI and Prime Inference (Muse Spark Open), CoreWeave…
- fix(config): fix bad config for providers or keys that cotains dots (self-defined)
- fix(scan): surface llmloop token-budget stop on the run-level state
- fix(frontend): update vulnerable development dependencies
- fix(gateway): scheduled loops, Slack DM sessions, and hosted coding agent on Fargate (7eadd2018) — Vaibhav Upreti
- fix: remediate GitHub code scanning finding #2417 - Updated [input_form. (b6fa04852) — Vaibhav Upreti
- build(gateway-image): ship the shell tools agents reach for (d60808365) — Vaibhav Upreti
- Nothing to do. The browser extension updates itself once no chat is busy.
- Projects on a server (SSH). Click + next to Projects, choose Folder on a server (SSH)…, and enter the server (for example me@server or a Host from your ~/.ssh/config) and the folder.
- Skills and files anywhere in the composer. / offers your installed Skills and @ the files and folders of the chat's project at any point in a message, not only at the start.
- 聊天侧栏置顶专家;业务流程专家;技能与技能包导出 ZIP
- 记忆页可手改用户档案与强记忆,冲突转入审核
- 技能、工具与团队的决策模型;通道插件;模型健康切换
Oct 16, 2026
- feat: add opt-in code push for GitHub git credentials
- feat: add worker capacity, agent backlog, and claim delay metrics
- feat: scale worker DB pools and support API-key model grants
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- One credits plan. Every account is on Free or Credits, with top-ups on either. Balances and usage read in credits everywhere; prices you pay stay in your currency. Raising your plan applies at once.
- Enterprise is a feature layer. SSO, SCIM, roles and groups, audit and branding come with an Enterprise license or a self-serve demo. Trials are gone. Admins can now manage billing.
- Restorable sessions. A deleted session can be restored for 30 days, with Undo on web and kortix sessions restore.
- Reasoning streamed as reasoning_content is sent back. Qwen, DeepSeek and GLM on their own
- Saving a provider in an installed Hive. The server checked API keys with code it imported
- release(langgraph): 1.2.16
- fix(langgraph): let other tasks finish when a resume value is invalid
- fix(langgraph): apply a resumed Command's update once
- Hosts can run Standard repository or path scans through the TypeScript runHostedScan(input, { executor, signal, onEvent }) API or the codex-security scan --host JSON-RPC adapter.
- Scan token totals include recorded usage from earlier Deep Scan worker attempts and context compaction without counting resumed history twice.
- Bulk scans preserve successful results and original scan failures when temporary-checkout cleanup fails.
- prepareQueuedMessages receives which queue (steering or followUp) a batch came from
- Fixed JSON Schema dereferencing rewriting literal values in const, enum, default, examples, and custom annotations, or treating property names as schema keywords
- Fixed Anthropic requests hanging forever on Bun 1.3 when the upstream accepted the request but never sent headers: aborting the request (including the first-event timeout) now fails it, so the…
- Your ChatGPT plan. The chatgpt provider runs TradingAgents on a ChatGPT Plus or Pro plan instead of an API key, through OpenAI's sign-in for open-source apps.
- What's new at start-up. The interactive CLI mentions a newer release, with the command that upgrades this install when it can tell how TradingAgents was installed, and notices from Tauric Research…
- Release downloads. Each GitHub release carries the source archive and a wheel, so a version installs with pip install from its release page without cloning.
- Add NVIDIA OpenShell as a sandbox provider.
- Show a banner when a prompt cache expires.
- Add /resume as an alias and improve how alias matches are ranked.
- Added CLAUDE_CODE_AUTO_MODE_MODEL to choose the model the client-side auto mode classifier uses; it does not change the model for server-side auto mode
- Added an Idle compaction row to /config for turning idle compaction off or back on; the notices shown after an idle compaction now say how to turn it off
- Added a notice to /design-sync and /design-login that standalone Claude Design (claude.ai/design) closes by December 14, 2026
- Connected apps on MausBot Cloud.
- Your Cloud now uses the connected-apps service included with your plan, with every app: Outlook, Gmail, LinkedIn, Calendly and the rest.
- Connect your apps once on your Cloud: connections made on your computer don't carry over.
- chat: check a reply's claims against the tools it ran (dcde468)
- chat: let people in a shared chat write in it live (fa4d503)
- chat: review quiet chats and fix what the replies left undone (cd19d1b)
- Usage ping: on each day you use Dossier, your server sends telemetry.dossier.aien.me a random install ID and the version number, once.
- The ID: made at random and kept in your database. It isn't derived from anything about you or your server, and exports leave it out.
- What's kept: the receiver keeps the ID, the day and the version for a year, and not the address the ping came from.
- Faster trace processing: traces are persisted and processed in fewer, larger operations for lower latency and fewer database writes.
- Code-graph docstrings: function/class docstrings are stored as graph documents and linked to code symbols for better code recall and search.
- New Memify API endpoints: convert content into memory entries via the API to make ingestion and downstream recall simpler.
- 8a5233b: Generate default chat IDs on first access so useChat() can prerender without reading a random value when the ID is not used during rendering.
- 894e130: fix(ai): add experimentallowMemory: true to streamText for single-consumer streaming without a replay buffer or retained final-content results, reducing memory retention.
- 0957554: Add box decision question type with image bounds for TypeSafe AI.
- pip install pr-agent no longer installs the git provider SDKs: PyGithub, python-gitlab, atlassian-python-api, the Azure DevOps SDKs, boto3, giteapy, the Google Cloud SDKs and a2a-sdk moved to the…
- Docker service images run as a non-root user: the webhook, app, polling and Mosaico targets run as UID/GID 10001 with HOME=/home/pragent.
- Azure DevOps and Bitbucket Server webhooks require credentials: with azure_devops_server.webhook_username/webhook_password or bitbucket_server.webhook_secret unset, every webhook is rejected with…
- First-run checklist: until your first session exists, the Focus view (and the Board on a phone) walks you through three steps: an agent CLI Glimmervoid can find, with the install command and a…
- Factory intents finish without you: a ticket that cannot be done inside its write scopes, blocks three times, or whose worker ends three turns without a change is now parked instead of stopping…
- A blocked test file can widen a ticket's scope: when a ticket's own change breaks a test file outside its write scopes, the orchestrator can replace it with a ticket that also covers that file, as…
- List remote agents for the mobile app.
- Bundle Kilo CLI v7.8.13.
- Prevent editor context ignore-file loading from blocking the JetBrains UI thread.
- feat(library): How Do You Embed an AI Agent Inside Your SaaS Product?
- feat(library): How Do You Build an AI Agent That Remembers Context Across Multiple Sessions?
- feat(library): 6 Best Microsoft Copilot Studio and Power Automate Alternatives for AI Agents in 2026
- The golden image of the inputs 0657d4ff61e9, built by the golden image workflow from ef69f5e19c1b46e41e743234fa5019e39442adcb.
- Fix: JS/TS class members differing only by _, $ or # are kept as separate nodes
- Fix: TypeScript this.field is typed from an assignment in any method, including static and accessors
- Fix: JS a workspace package whose exports points only into dist/ falls back to src/
- Power and memory: Macs can idle-sleep again while the Claude CLI source is in use, because passive probes no longer leave Claude Code's caffeinate running.
- Claude: optionally read usage from the Claude desktop app's own signed-in session; map each token account to its own Claude config directory for local cost.
- Providers: an explicit Gemini app source for the 5-hour and weekly limits, Munito and Future AGI usage, Antigravity Pro accounts no longer labeled Free, Claude Cloud credits in widgets, OpenCode…
- Full Changelog
- Adds independent confirmation to the opt-in contrastive API. A development improvement now has a separate path for testing on unseen task pairs before a caller considers adoption.
- 1. Frozen candidate, baseline, environment, execution envelope and decision policy. 2. Exact paired statistical comparison, minimum gain and regression gates, with every declared case required. 3.
- Measured pilot: development improved 23/24 to 24/24, but baseline, contrastive candidate and control all scored 38/40 on confirmation. Both superiority tests rejected with p = 1.
- Native companion binary release for oh-my-opencode-slim.
- AIgateway declarative provider
- Propagate W3C trace context from ACP prompts to MCP tool calls
- Surface provider x-request-id on responses and errors
- #957 b1646e2 Thanks @zichen0116! - The agent's tap, double_tap, and right_click tools, in agent.act and the MCP server, take an optional modifiers list (Shift, Control, Alt, Meta, ControlOrMeta)…
- #1192 40bea9e Thanks @adc77! - Report the model behind cached actions and identify replays made with a different model.
- #1029 7d067af Thanks @rudycelekli! - Align failure code-frame carets and clip source lines by terminal display width.
- @Enard52 for the KiloCode link repair and hidden-adapter discovery fixtures.
- @Amiirhosseini for the contributor setup guide, the tracked-inventory proposal, and the overlapping KiloCode link repair.
- feat(web): link personal feature previews from organization settings
- feat(traces): integrate PostHog analytics for IO parser comparison tracking
- feat(skills): improve browsing and editing UX
- c1b079c: Publish framework-neutral tracing as @vercel/agent-runs with singular createAgentRun and AgentRun APIs; eve still bundles it without a new runtime dependency.
- Bots see what their teammates are doing: list_bots says what each other Bot is doing now: waiting for your approval, moving to a fresh conversation, working…
- Hand-offs that come back: a Bot that hands part of a request on to other Bots, or posts in a group and asks for the replies, answers the Bot that asked only once all of it is back, and its answer…
- A Bot's character shows when it waits for teammates to hand back work: it sways with its eyes circling, without the working dot, and its sidebar row says whom it waits for.
- Sandboxed Harness Agents: ClaudeAgent and CodexAgent take a sandbox= parameter and run the whole harness process inside an isolated, per-session computer instead of on the host, with session state…
- Docker Sandbox: DockerSandbox is the customer-operated provider. Containers run as UID 1000 with a read-only image filesystem, dropped capabilities, no privilege escalation, CPU, memory and…
- Amazon Bedrock AgentCore Sandbox: AgentCoreSandbox implements the same contract on AgentCore Runtime, with AgentCoreCheckpointTransfer next to the Docker transfer.
- Show agents that talk to each other on the website
- Name Synthetify Labs as the licensor and add a commercial contact
- Write only the changed mailbox rows on each mailbox change
- [client-v2] Fix swipe-back / browser back exiting the app after link action navigation in the v2 (mobile) client
- Configure model reasoning options to generate reasoning variants, replacing catalog variants while preserving your configured variant overrides.
- npx claude-mem teambrain join <invite-link> connects this install to a TeamBrain team using the same device-code sign-in as install.
- Members who join with a team seat keep their own local observer — joining never switches your observer provider.
- When an install belongs to a team, SessionStart adds a short team brief: open collisions in this project, recent plans and decisions, who else is active, and items delegated to you.
- #2931 factory daemon start no longer leaves factoryd unloaded (launchctl EIO) when it is already running
- #3050 Shared probe detects the target repo's lint config
- #3005 SHIP skips the CI wait sooner when a repo has no CI configured
- Windows 新建或恢复会话时反复弹出黑色终端窗口(#1593,#1583):根因是桌面 sidecar 被打包成了 GUI 子系统,它自己和它拉起的控制台子进程(例如每次启动 CLI 都会跑的机器码探测)因此各开一个可见窗口。v0.7.3 只处理了其中一条路径,本次把 sidecar 恢复成控制台程序,并在构建时检查子系统类型,避免再次回退。
- + 菜单一级直达 Agents、工作流和斜杠命令(#1601,#1597):运行一个 Agent 或工作流不再需要先展开「更多工具」,一次悬停或点击就能到;「更多工具」这一行已移除,手机底部面板同样少一级。
- Linux 包的终端面板(#1602,#1591):此前的 Linux 包没有带上 node-pty 的原生模块,终端面板一律报「Failed to load native module: pty.node」,现在安装包会正确编译并校验该模块;同时把 Linux 模块改为按 glibc 2.28 编译,Ubuntu 20.04、Debian 11、RHEL 8 等环境也能正常使用终端。
- Finished topics get put away, not deleted. A stretch of conversation that has ended was the most
- It reads a detail back instead of guessing. Ask about something a summary left out and it opens
- A choice card's question and answer stay in the conversation. The card used to come and go, so
- Media: Upscale, Remove background, Generate captions and the new Remove silence run on the selected clips. If nothing fitting is selected, a bar over the canvas asks you to pick a clip first.
- Export: choose the resolution for a video or the current frame right in the list.
- Assets: move the selected media into a folder, or reveal it in the assets panel (⇧⌘O) or in Finder / Explorer (⌥⌘O).
- Import FreeCAD and Rhino files (.FCStd, .3dm) through cadmpeg, and glTF 2.0 (.glb, .gltf) as mesh bodies with their colours and node structure.
- Joints: gears, rack and pinion and screws; inverse kinematics with joint.reach, and a jointed part follows a drag; Motion Trace keeps a point's path as a joint is driven.
- Sketch: associative pattern tools with Flip/Center; infinite construction lines; a new sketch on an origin plane faces the side you clicked; inch and foot marks in expressions.
- [DND-1779] [INFRA] chore: move dependency updates from Dependabot to central Renovate
- [OPIK-8777] [SDK] fix: track ADK agent spans opened via start_span on google-adk 2.12.0
- [OPIK-8766] [BE] feat: return the latest prompt version when a save does not change it
- 640788c059537cd60daf11d152bd76a2e3fc2a9f MUL-6945 & MUL-6944: fix(daemon): preserve Git content during artifact GC
- da1671ca204e229cba6878a99f21942b5ad9af32 MUL-7420: Project workflows — columns and handoffs on the workflow
- 700d25eef51b020dce408a75cf0b2d1378851dac MUL-7420: render the workflow brief once on the server, for the assignee or an observer
- skill 仅增加一段简洁说明,明确观察截图不会保存到手机相册,避免转到 Mac 截图、剪贴板或网页传图。
- 新工具复用共享操作锁、锁屏检查和认证暂停保护;不自动重放不确定的截屏请求。工具总数为 18 个模型工具和 2 个 widget 工具,保持现有 schema 体积限制。
- 同步插件与 widget 版本,以及工具数量相关的打包、文档和宿主适配检查。
- Durable libfx sessions can continue after crashes, redeploys, and function timeouts. Interrupted tools run again only when marked idempotent: true.
- You can run multiple named libfx agents in one app, each with its own model, instructions, and tools. Sessions stay with the agent that created them.
- wakeHandler() provides an HTTP handler for queued prompts and recovery, so one Vercel route serves all your libfx agents.
- The PageIndex SDK, local or cloud — vectorless, reasoning-based RAG, end to end.
- Much faster indexing — the PageIndex Flash engine gets the tree from layout stats: no LLM involved for the structure generation itself, LLMs only write the node summaries, and tree expansion…
- Flash engine: the local default (mode="standard" keeps the classic LLM pipeline).
- buildInputPreview() and INPUT_PREVIEW_MAX_LENGTH are deprecated in favor of buildPreview() and PREVIEW_MAX_LENGTH.
- Factory sessions no longer have modes (Plan/Explore removed); thinking defaults and related config endpoints now use a single global default, and PUT /web/config/thinking rejects modeDefaults.
- Factory credential behavior changed: cross-scope credential fallback is removed; custom credential providers must move from orgFirst to explicit credentialScope / policy resolution.
- core: Suspend in-flight production executions at worker shutdown (5863179)
- Marketplace plugins such as goal load again. In 0.0.46 they failed with Cannot find module '@cline/sdk' because the packaged app didn't include the SDK packages plugins import
- The Cloud sessions toggle is gone from Settings. Cloud sessions are on wherever they are available
- When a provider error interrupts a response mid-stream, Cline retries it up to 3 times before failing the turn
- 原生 worker 与运行时扩展到 Linux x64:以 x86_64-unknown-linux-gnu 为一等目标构建、测试与打包,npm run build:ptc 同时 stage Linux worker 及其独立 receipt,POSIX 目录 fsync 与 Windows MoveFileExW 发布按平台分别实现。 @Thatgfsj, @T-Auto
- 新增 macOS Apple Silicon(darwin-arm64)支持:随包提供带版本、目标、可执行位与 SHA-256 receipt 的 Mach-O worker。 @null-characters
- 文件工具在解析操作路径时规范化父链,使 /tmp、/var 等系统别名下的路径可用,而直接命名符号链接的最后一段仍然 fail-closed。 @liuchunyi-buaa
No release notes published.
- Added terminal status reporting over OSC 7501: in terminals that support it, the tab or window shows whether the CLI is working, waiting for your input, done, or failed.
- Added the alt+m shortcut (Option+m on macOS when Option is set to act as Alt) to send a message using the opposite of your Queued Messages setting in /config, so you can send right away while in…
- Improved agy plugin validate so it checks a plugin with the same rules the CLI uses to load it, instead of simplified checks that rejected valid plugins (such as a manifest without a name) and…
- commit: b67a1ae613c3449a9f33635d08481e65de59fe9b
- source tag: beta2026-10-16_11-38-16
- regions: cn global
- Redesigned Automations — Start from six example workflows, see upcoming runs, and inspect results in a dedicated run dashboard. Pause, resume, or request a run from the automation list.
- Unified model configuration — Search configured models, manage provider configurations, and choose a default for new Sessions from Settings → Models.
- Work roles in your profile — Choose your role during onboarding or skip the step, then update it later in General settings.
- Native chats have their own permission mode. Choose permissions for each new Claude or Codex chat from Settings → Chat → Behavior.
- Native chat sessions are easier to continue. You can resume large terminal conversations in a new native chat, recover chats whose only work ran in the background after a restart, and see every…
- Claude accounts and Codex chat setup are smoother. Claude accounts again use separate folders, the sign-in link can be copied, Codex chats require version 0.136.0 or later, and Codex Ultrafast is…
- wren: add StarRocks as a first-class data source (4e1c25e)
- context: count errors in the validate warning summary (fedf182)
- context: keep cubes when importing an MDL with init --from-mdl (e65351e)
- Attach namespaced origin and tool_call_id metadata to BackgroundTools completions
- Take GPT-Live realtime history, usage, and reply waits from the session core
- Honour disable-model-invocation and user-invocable in Skills, and skip stray skill library paths
- Package ego-browser plugins for nine agent hosts
- 2.0.0 beta dev
- fix(plugins): make the Codex archive uploadable to the public plugin directory
- server: export server logs over OTLP (add77bf)
- playground: forward Bedrock tool results directly as text (15355fd)
- add Atomic Agents tracing integration (fe5e933)
- feat: AG-UI 1.0 for CopilotKit (7693a04)
- fix(runtime): ping Intelligence sockets every 15s so 30s proxy idle timeouts don't drop runs (75f5721)
- fix(channels-slack): deliver a message click's state.values to the handler (269fe33)
- fix(slack): keep the task plan off Slack replies (27d85b3c7) — vaibhav upreti
- The Gateway keeps running after the computer sleeps or the system clock jumps, instead of restarting itself because its ownership heartbeat looked stale.
- Gateway shutdown and restart no longer hang on a stalled transcript worker or in-flight session lifecycle events. Thanks @Oldrich333.
- Tlon group DMs no longer treat other participants as the owner, closing an owner-impersonation path. Thanks @eleqtrizit.
- CLI archives built from the Studio release studio-v2.42.0.
Oct 15, 2026
- Handle temporary empty Cloud tab URLs during navigation.
- Keep screenshots, crops, and pointer coordinates aligned on scaled displays.
- Provide 18 default browser tools and six optional actions through Browser Use Actor.
- feat(runtime): assign Trajectories to Learning Spaces on the server (81ca993)
- fix(web-inspector): coalesce streaming transcript refreshes (16a1b46)
- fix(a2ui-renderer): type RendererProps.children as (id, basePath?) (5bdb483)
No release notes published.
- Agent chat on Windows: agents installed with npm (Claude Code, Codex, Gemini CLI, OpenCode) failed to start with "%1 is not a valid Win32 application (os error 193)".
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Open /config to search preferences across categories, change supported settings immediately, and see which changes require a restart.
- Send a follow-up while Codex is responding to steer it immediately. Interrupting a turn also stops active Code Mode cells and their nested tool calls.
- Configure a leader key in /keymap. The default is Ctrl+X, shortcut hints stay open until you choose or cancel, and voice mute moves to F9.
- 4210766: git({ github }) accepts an optional authorize callback that runs once per gh tool invocation with the caller, declared repository, and repository-scoped token.
- 3fea287: Dropped traces and policies excluding either content direction now request Workflow data deletion when the run ends.
- a5e33ec: Root sessions now accept empty model responses with finishReason: "stop" without retrying them or reporting a model-call failure.
- Run on real devices from kane-cli run or the TUI — real devices are selectable in both, and a live device window shows the session as it happens.
- Real-device runs work from any machine — no local device setup required, and --remote now picks sensible device defaults.
- You're told exactly which plan a real device needs — when the hub refuses a request on plan grounds, the message names the required plan instead of a generic failure.
- fix(bedrock): drop reasoning blocks the target model rejects when history crosses model families (5cb880559)
- fix(subagent): surface the child's stop reason instead of masking blocked runs as success (ab88efd81)
- fix(context): keep removal markers and summaries out of message 0 (47fb500af)
- Updated bundled Claude CLI to version 2.1.301
- Async subagents now report when they’re waiting for a human.
- Async subagent runs now appear under their parent’s trace and identify the agent that launched them.
- Summarization middleware now exposes its summarizer.
- Added CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS so unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG) ends a request at a usage limit that resets too far away instead of waiting
- Added OpenTelemetry tracing guidance to the claude-api skill, including setting up trace destinations with the ant CLI
- Added the % motion to vim mode: jump to the matching bracket, alone or with d/c/y
- Session sharing has been discontinued. Existing share links and session-sharing commands are no longer available, and importing sessions from share URLs is unsupported.
- Choose where /share uploads: sessionShareDestination selects auto, gist, or radius. See Sessions and context and Export or share a session.
- Cache-preserving thinking level changes on OpenAI GPT-6: changing the thinking level mid-session with an API key keeps the prompt cache.
- Kimi Code over Chat Completions: tools added mid-conversation no longer invalidate the prompt cache. models.json base URL overrides must now include /v1. See Providers.
- Show and edit agent models from the agents list.
- Configure transcript display, with a live preview and localized settings.
- Reuse the Core submenu in the prompt bar menu.
- Nothing to do. The browser extension updates itself once no chat is busy.
- Fresh turns for long work. In 2.1.38 the automatic continue often couldn't be sent.
- No chats in incognito windows. If the extension is allowed in incognito and an incognito window was in front, a chat the app opened (for example a sub-agent's) landed there.
- Fixed Hoplite showing an error in Settings → Providers instead of discovering models in workspaces linked to a Hoplite project
- HighLevel (GoHighLevel, Growably and other white-labels). Open Connected apps, choose Connect on the HighLevel card, sign in and pick your location.
- Files. A new Files page in the sidebar shares a folder from this computer with My Cloud, so your bots there can read it.
- Use your own Claude Code skills. A switch lets Claude bots use the skills you already have in Claude Code, in place.
- 739e9f6: Restore raw OpenAI custom-tool input deltas for consumers that forward them to Responses API clients. Mark streamed input with inputEncoding: 'text' so UI clients can display it…
- Updated dependencies [739e9f6]
- @ai-sdk/provider@3.0.19
- Added Claude Opus 5.5 and Claude Sonnet 5.5 for Application Default Credentials (AGY_ADC_AUTH) sign-in. Claude models now appear in /model only if your Google Cloud project can use them.
- Improved Auto mode's review of commands that would leave the terminal sandbox: the review is faster and no longer times out without a decision, it allows the workspace's own format, lint, and fix…
- Improved the /diff Commit Tree view with colors: commit hashes and change IDs are highlighted, descriptions are dimmed, the Working Copy (Uncommitted Changes) row stands out, and @ is bold, so…
- 修复 Windows 桌面端打开应用、新建或恢复会话时,可能反复弹出黑色终端窗口的问题(#1585)。
- Memory: provider plugins no longer accumulate memory on every refresh. QuickJS worker temporaries are released and retired JavaScriptCore runtimes are freed, so a 20-provider synthetic workload…
- Providers: Claude Platform shows the monthly Console API credits included with Max and Team plans; Muse Code can earn one explained Keychain grant on a manual refresh.
- Settings and safety: a config file that fails to load is never overwritten, and About now lists open-source acknowledgements with the full license and notice texts, which also ship with the CLI…
- fix(managed-agent): connector and broker robustness from the #12692 R2 review
- feat(managed-agent): H4e-b2 the lead's mailbox over Session messages
- fix(managed-agent): settle worktree child runs after their merge (#13753 I2 follow-up)
- Read the source turn of mailbox messages once per lookup
- Answer approvals from a notification and a paired Apple Watch
- Center the icons in message chips and give them one size
- CDP. The new browser control is rebuilt with chrome.debugger API, it uses CDP to interact with elements in web pages. It's more stable than the previous injected scripts solution.
- One agent, one chat per tab. The old planner and navigator are replaced by a single agent. Several tasks can run side by side, each in its own tab group.
- You decide what it may do. Ask mode (default) asks before the agent submits, sends, deletes or uploads; Auto mode goes ahead. Payments, passwords, codes and access grants are always handed to you.
- bb desktop 0.46.1
- Tokenizer.countLatestTurn: counts a message as the transcript's latest assistant turn, which an Anthropic Messages replay keeps byte-for-byte
- Agent.prioritizeSteering: moves given pending steering messages ahead of older ones so one-at-a-time delivery takes them first
- Fixed a Cursor tool call re-sent by the server persisting two results under one id; the re-run's result replaces the first
- chore: set graphiti-core version to 0.31.0 for the v0.31.0 release
- chore: set graphiti-core version to 0.31.1
- every Sandbox daemon opens one shared Store directly;
- each mounted Workspace has its own overlay;
- Commit merges with a Branch head that has moved since the mount, and on
- fix: use provider token counts and recover from context overflow
- fix: stop claiming after cancellation and harden live provider tests
- perf: scale agent execution with typed execution state and a claim dispatcher
- Conversation view (experimental): turn on Conversation view in Settings > Sessions, then pick Conversation view in the restart menu of any Claude Code or Codex card to try it on that session alone.
- Streamer mode: a mode for screen sharing and live streams. While it is on, tokens and keys…
- Session cards show their pull requests: a session's card header and its Focus roster pill now show the PRs it works on as #number chips.
- [OPIK-8748] [BE] feat: a model catalog with profiles and per-model params, served by the registry API
- [OPIK-8706] [BE] Bound trace-keyed spans reads with the span_weeks index
- [OPIK-8329] [BE][FE] fix: offer Ollie widget building only in Custom Charts workspaces
- The golden image of the inputs a92891c4ded8, built by the golden image workflow from ef69f5e19c1b46e41e743234fa5019e39442adcb.
- #1139 b292019 Thanks @Anmol202005! - The bug-bash skill's config example removes its repro-directory exclusion before adding the repro glob, so projects that exclude tests/bugbash/ from regular…
- #1149 6d686f1 Thanks @okwasniewski! - Piping CLI output into a reader that stops early, such as e2e --help | head -1, no longer crashes with an EPIPE stack trace.
- #1150 93bf5d8 Thanks @okwasniewski! - An openHost() session closed after its deadline now always ends as timed out: close() returns the TEST_TIMEOUT and the trace says timed out, even when the…
- Docs: a Convert menu on the Home tab and in the File menu; a picture inserted into a table cell stays in the cell (reported
- Sheets: conditional-format formulas using ROW()/COLUMN() are 1-based, so header fills no longer vanish (reported
- Slides: chart title and legend geometry measured against PowerPoint, chart userShapes overlays, rich data labels, small caps, hidden axis lines and line legend keys.
- Run approved shell commands with Bash when available and POSIX sh on
- After installing on Linux, run the packaged runtime's host setup check and
- Bundle Astrid 2026.10.1 with its native runtime-media isolation, Linux
- Safer September upgrades: updates on the 2026.9 line keep state, plugin data, service ownership, and Windows state paths intact, and refuse automatic restore over unattributed Doctor writes.
- Updates that finish on pnpm, npm, and Windows: pnpm 12 Gateway services and launchers survive native updates, retained updater runtimes no longer block npm recovery, and Windows updates no longer…
- New model support: run Meta Muse Wonder, use MiniMax M3.1 thinking effort levels, and stay signed in to MiniMax portal OAuth past token expiry.
- feat(llm): add organization-level connections
- feat(billing): record billable units in an append-only ClickHouse ledger
- feat(skills): add CLI import instructions
- A revision is a new version of a page, not a second page. Ask for a change and the panel keeps
- The brain can see the pages it made. list_pages and get_page are two new read-only tools,
- A fresh install comes up light. A device that has chosen a theme keeps it; the dark palette is
- agent-framework-azure-cosmos-memory: Add a dedicated Cosmos memory provider User-Agent token
- agent-framework-foundry: Surface OAuth consent requests in non-streaming responses
- agent-framework-ollama, agent-framework-openai: Add configurable service URLs to chat clients
- The page_number metadata field now means the same thing in every splitter: the page the chunk's text starts on. Two cases change
- RecursiveDocumentSplitter reported the page a chunk ended on. A chunk spanning a page break now reports the page it starts on, matching DocumentSplitter and MarkdownHeaderSplitter.
- DocumentSplitter, MarkdownHeaderSplitter and EmbeddingBasedDocumentSplitter reported the previous page for a chunk that began with one or more form feeds (\f), even though all of the chunk's text…
- feat(project-files): add public sharing and protected delivery
- feat(project-files): add cross-owner copy and bulk transfers
- feat(project-files): add rendered delivery and indexed search
- Add harness updates, logout and model defaults
- Make session switching instant with no layout shift
- Use official harness logos with light and dark variants
- The gateway starts after a reboot again. After a restart, Linux can hand an early PID to an unrelated process, and the gateway could mistake it for a running gateway, so a supervised gateway…
- Long local-model turns survive on laptops. Local requests always stream with prompt-reading progress (shown in the status line), stall timers skip time the machine was asleep, a recovered local…
- Release: split the Microsoft Store bundle into its own windows-store lane (f9d8a740) — @ethernet8023
- ROB-1291 Multi-tenant Prometheus toolset (Mimir X-Scope-OrgID routed by a Kubernetes label)
- ROB-1525 Reuse the server's cached tool executor in /api/info and the OAuth callback
- Keep an executor rebuilt while a toolset refresh runs
- Knowledge Base: Special characters such as \, /, ?, *, and [] are now disallowed when renaming documents.
- Agent: A confirmation prompt is now shown when deleting conversation history, to prevent accidental deletion.
- Q&A Page: Tables in AI responses now support horizontal scrolling when the number of columns exceeds the visible width.
- core: Create an XML parser for each webhook request body (ea3ee53)
- pi-web-access 0.41.0. Sites that turn away AI fetchers can be read by setting your own User-Agent: add "fetchContent": { "userAgent": "..." } to ~/.feynman/agent/web-search.json.
- Add six theme palettes and custom colors and corner radii, with live preview and saved preferences
- Add an authenticated, multi-tab terminal inside Chat, with a working directory for each conversation
- Allow Console font size to be adjusted from 12 to 20 pixels across chat and settings
- playground: add Meta muse-wonder-1.0 and gemini-4-argon (d5168d1)
- oauth2: preserve SameSite=Strict cookies on server-owned return_url after IdP sign-in (c332b0c)
- add Claude Managed Agents tracing docs (54fe899)
- MCP initialize、ping、工具与资源发现不再上报启动。仅首次实际工具调用结束时记录该进程的使用启动。
- READY、setup 和预览结果字段合并到单条 iphone_use_tool_called,停止重复的专用结果事件;首次使用仍保留一条独立的使用启动事件。
- pua_screen_action(action="viewport") 自动几何恢复与失败均不埋点;帧轮询继续排除。用户预览点击、拖动与工具栏操作正常记录一次。
- OpenAI GPT-Live 1: now available through LiveKit Inference.
- Google Gemini Live: adds Inference model-string support for google/gemini-3.8-live, google/gemini-3.8-live-extended-thinking, and google/gemini-live.
- xAI Grok Voice 3.0: replaces Think Fast 2.0 in the Inference and xAI plugin model lists.
- commit: 0b779c8210fa4b9363fb1db8b8d7e5fab46019d4
- source tag: beta2026-10-15_12-33-25
- regions: cn global
- uses: openai/codex-security@3f61e5150981a5087c32e4395bbaa8ce5ce0c3e3 # action-v0.3.0
- The shopify_admin provider has been removed and consolidated into shopify. Update service references and configure a connection under shopify; there is no compatibility alias for the old service name.
- For API-key connections, use the Shopify Admin API access token and the shopDomain field (for example, acme.myshopify.com). OAuth client configuration uses shopSubdomain (for example, acme).
- Re-read each action's input and output schemas before migrating. GraphQL actions use string global IDs such as gid://shopify/Product/123, cursor inputs such as first and after, and pageInfo…
- fix(geometry): unify cross-frame coordinates and screenshot conversion | 统一跨 frame 坐标与截图转换
- refactor(observation): decouple DOM processing from capture orchestration | 解耦 DOM 处理与采集流程
- refactor(observation): unify cross-frame capture coordination and data ownership | 统一跨 frame 采集协调与数据归属
- Jira and workspaces: Jira now has boards, sprint and status views, saved filters, planning columns, reusable prompt templates, quicker assignment, and safer agent access from the CLI. Folder…
- Native chat and agents: Native chat can show inline edit diffs, wrap long code lines, zoom Mermaid diagrams, display Codex context usage, and retain the correct permission mode. More terminal…
- Windows and WSL: Orca can run an experimental managed host in a WSL distro, with clearer recovery and account handling. It also improves Windows hooks, daemon checks, and the Windows-to-WSL path…
- Update custom error formatters. RunConfig.tool_error_formatter can receive kind="tool_exception" and an optional error. Handle the new kind or return None to retain the default message.
- Plan upgrades around saved runs. New RunState snapshots use schema 1.21. Finish affected older same-name agent graphs on their original SDK. Keep the agent code and original hook ordering unchanged.
- Recover completed work more reliably. Deferred approval history, completed tool outputs, aliases, and compaction state survive supported recovery paths. Automatic compaction waits for approvals.
- wasm: expose cube orderBy in TypeScript SDK types (af1269f)
- bindings: build against wren-core 0.3.2 (2b0543d)
- chore: remove unused helpers (b5092284d) — vaibhav upreti
- feat(shell): pick exact tasks for cron logs, run, and remove (deb0df718) — Anwesh
- feat(shell): edit goals, work items and loops with prefilled forms (e60300584) — Anwesh
- Settings → Model preferences gains an independent naming model and per-role reasoning effort, so titles can use their own model and each role its own effort. Refs #11851.
- New side questions: ask a read-only question beside the main conversation; read-only tools, never in the main transcript, dropped on close. Refs #12927, #13128.
- Each class in the context details can be opened to its largest items (sizes and tool names only, never message text). Refs #13052, #13143.
- Reduce idle CPU and disk reads in repositories with many untracked files
- Share model catalog processing across equivalent clients while keeping usage and account state fresh
- Allow Desktop-capable clients to receive Markdown responses through the new user_response tool
Oct 14, 2026
- Agent): add nemotron super 3.5 default parameters (c59d2994) — @elisam0
- Cli): add curated model list at the top in the model picker list (e6835fd3) — @elisam0
- Plugin-catalog): add hermes-lang-ko — Korean Desktop language pack (e5f49b6a) — @dandacompany
- Publish zepai/graphiti on v*.. tags like MCP
- Add pevans to the CLA allowlist [ZEPAI-3570]
- Update root uv.lock to fix Dependabot alerts [ZEPAI-3570]
- 对话 mid-turn steering
- 当前专家记忆备份、恢复与存储维护
- Telegram 可限制允许的用户 ID
- Suppress ownership warnings for root installs.
- Avoid restyling transcript history when inserting messages.
- 4bfa520 release(deepagents-code): 0.1.85
- antigravity: add native Google Antigravity IDE integration (cd4367b)
- opencode: support the OpenCode 2.x plugin API (1d4b357)
- plugins: add headroom-snip Claude Code mod that animates compression (17abee2)
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Fix: MCP graph_stats no longer hangs under a stdio host on Windows (and no longer silently drops its built-at-commit staleness line): _git_head closes the git child's stdin so git does not block…
- Fix: Rust a use crate::…::Name edge now lands on the declaration it imports rather than a bare name-folded target
- Fix: JS/TS a file-scope callback's own parameters are now shadowed, so a bare parameter passed on as a call argument no longer mints phantom cross-bundle indirect_call edges
- Removed ApprovalDetailsOptions; formatApprovalDetails takes only args again, since the compact approval prompt no longer shows tool details
- AgentTool.delegationGuidance: instructions a same-name extension replacement inherits in its description when it delegates to the native tool
- CompactionPreparation.lastSummarizedEntryId and CompactionResult.providerReplayThroughEntryId: a remote compaction that could only fit the history before the kept tail reports the last entry its…
- SolidCraft closes right after launch on some Windows PCs. On Windows SolidCraft now starts on Direct3D 12 first and only tries Vulkan, then OpenGL, if Direct3D 12 has no GPU.
- You can still choose: set WGPU_BACKEND to dx12, vulkan or gl, or put "backend": "dx12" | "vulkan" | "gl" in graphics.json in your settings folder.
- A new CI check starts the real app on Windows in six setups (default, MSAA, a damaged settings file, safe mode and others) and fails the build unless each one draws and quits cleanly.
- briefing: approve a proposal card on the server (99eca81)
- briefing: open the first week's briefs with getting-started moves (1a28256)
- chat: pin, rename, share, export and redo the last turn (8068e43)
- Bump outdated patch/minor dependencies
- Re-prompt the planner with validation errors when the design artifact is invalid
- Explain what the run loses when the design artifact stays invalid
- Eyes: the iPhone mirrors its screen over AirPlay to UxPlay on a Mac or Linux computer.
- Hands: a ~$7 ESP32-S3 plugs into the iPhone's USB port as a mouse + keyboard (AssistiveTouch pointer), connected to the computer over Wi-Fi.
- Brain: any OpenAI-compatible vision model (demo: Qwen3.8-Flash-Next on two DGX Sparks).
- feat(skills): package Strands for Claude, Codex, and Cursor (5e84cafc7)
- feat(ts): add DecisionModel primitive and LLMDecisionModel (8340c4241)
- feat(mcp-ts): add progress callbacks for tool calls (5eaf888f6)
- improvement: [All SDKs] update bundled Copilot CLI dependency to 1.0.98-2 (no SDK API changes)
- Add support for Gemini 4 Argon
- Command hooks complete reliably even when child-exit signals are lost under Node
- MCP status and /mcp commands stay synced to the active session after split-view session switches
- 68b9d58: Approved defineTool calls can read who approved them: ctx.approval.responder is the approver and ctx.approval.responder.getToken(provider) resolves their token, while the call still runs…
- 620022b: The Braintrust eval reporter no longer aborts the run when an eval produces no agent output. Those evals are now logged with an empty output.
- 2e45423: Allow eve/client to use an AbortSignal that does not implement throwIfAborted, including the signal provided by React Native.
- Updated bundled Claude CLI to version 2.1.300
- Give Artifacts its own tab with Finder/Terminal shortcuts
- Let a global default profile fill in for unbound sessions
- Show the board sooner on launch
- feat: add machine pool connection test and abandon stuck BYO deletions
- feat: make integration-launched agent configs editable
- feat: Agent config json definition + launch time config
- 1ab3430a4b6c8add9f08854c77544b0211a3f13c .NET: Dotnet: Add approval reuse considerations
- 563eb3a9965f7c144ad697223320411c02fa349a .NET: terminate chat completions SSE stream with data: [DONE] [ #8526 ]
- 4964bdd1b2d55b8137cbc1bc85840181af05bda0 .NET: Update version for 1.25.0 release
- SSH and managed servers: New SSH connections now use Orca's managed server. Terminals left running by v1.4.224 keep working in their existing tabs through the upgrade and reconnect; new terminals…
- Native chat (experimental): Code blocks and shell commands are highlighted, attached files show as filename pills, and you can edit a queued message even while the agent is asking a question.…
- Agents and accounts: Kiro reports agent status and appears in Session History. Claude accounts use separate folders and offer Copy sign-in link instead. Orchestration workers can start as native…
- Added virtualText to a mod's prompt decorations: text drawn in the prompt box that never enters the draft or the submitted prompt
- Added cache_creation_5m_tokens and cache_creation_1h_tokens to the OpenTelemetry api_request event, splitting cache writes by lifetime
- Added $.ui.view({ agentId }) for mods: shows an agent's transcript in the main view when you ask, and $.ui.view({}) goes back
- #40 be758b9 Thanks @PazChen! - Desktop: an agent's own notices show apart from its reply.
- #39 9c05395 Thanks @youyouhdhd! - Settings → Agents: an agent found and connected has Locate… too, beside Disconnect, so the program it runs can be chosen even when one was found — a particular…
- #46 e4e67fe Thanks @PazChen! - Studio: a title picked on the picture can be restyled together with every title like it.
- fix(desktop): Cmd+; only offers cloud workspaces that are in your sidebar
- fix(mobile): chat feedback pass: keyboard dismiss, folded tool calls, glass header, dock chips
- feat(desktop): chat links follow the link click settings, Codex-style
- Changes since langchain==1.4.5
- release(langchain): 1.4.6 feat(langchain): add opt-in per-call interrupts to HITL middleware
- #1132 c82a03a Thanks @okwasniewski! - app.command.inheritEnv passes host variables to the app by name or trailing-* prefix (['DATABASE_URL', 'STRIPE_*']), copying each one only when it is set, so…
- #1134 8497c06 Thanks @okwasniewski! - E2E_CACHE_DIR and e2e run --cache-dir <dir> (also on e2e cache) replace cache.dir, so checkouts and worktrees of one project can share a replay cache without…
- #1137 27a6ed0 Thanks @okwasniewski! - e2e feedback now asks for reports the e2e team can act on. A --type bug report needs --repro, steps that reproduce it every time.
- Improved Agent Panel terminal support by allowing Terminal tabs to be dragged into it. (#60582; thanks cppcoffee)
- Improved custom Amazon Bedrock model support with configurable tool use, images, and thinking. (#64953; thanks mpyuglgwkxcmodyrpo)
- Improved memory and CPU usage while the agent responds in long threads.
- ruflo tui draws the console's pictures, looks and motion (console 0.44.x, #4182, #4188): --look bbs|orbital|plain, --fps, --anim, --pictures, --boot, and any console view with --view.
- ruflo init: the v3 implementation skills are opt-in (--full adds them), like agents.v3 (#4140, #4139, @z4forzubair).
- MCP server and daemons: the server starts in the session project and ignores a CLAUDE_FLOW_CWD that names the home directory or a filesystem root (cli-core 3.7.3).
- Group permissions: a group chat can run all its members with one permission preset (Read only, Workspace or Full access) in place of each member's own.
- A Bot at work in a group chat no longer stays silent until it is done: it says what it will do, keeps the group posted as it goes, and ends with the result.
- The work status in the group, the sidebar and a Bot's own chat says what the Bot is doing, instead of only "On it", also while it moves to a fresh conversation.
- ui: keep labels on Escape and fix table sort, a11y and accessor warnings (53a1735)
- Prefix shortcuts work again when the terminal sends the next key as plain text. This affected Tern, Chinese, Japanese and Korean input methods on macOS, and Ghostty text: keybinds that send the…
- Codex sessions no longer start with a "clamping Interrupt hook timeout" warning. Herdr no longer installs Codex turn hooks, and Codex status comes from screen detection again.
- Claude Code questions show as blocked in named sessions.
- #4159 0a3a2fb Thanks @RealKai42! - Add an opt-in NotifyUser progress update feature (off by default): new sessions can show the agent's mid-turn progress with an optional one-line title, enabled…
- #4145 242ac23 Thanks @wbxl2000! - Add kimi app [path] to open a new desktop chat in the current or specified directory.
- #4061 20a2cea Thanks @chengluyu! - Cap the agent's in-turn wait for background tasks at 90 seconds (longer timeouts are no longer accepted), discourage repeated waits, and let a new message end…
- Put message actions in one row under each message, and add reactions and text selection to the mobile app
- Add Vercel to the Apps catalog
- Let agents capture the screen after the user approves each capture
- No more freezes on busy workspaces. In 0.1.108, a workspace with many finished team jobs could keep the bot server at full CPU until it stopped answering.
- OpenMausBot starts on Windows even with a package.json above it. Sometimes a folder above the app (usually your user folder) had a package.json saying "type": "commonjs"; npm init -y makes one.
- Bots start with Full access, on every engine. Like other agent apps, every new bot and conversation now runs commands and edits without stopping to ask.
- fix(geometry): unify cross-frame coordinates and screenshot conversion | 统一跨 frame 坐标与截图转换
- refactor(observation): decouple DOM processing from capture orchestration | 解耦 DOM 处理与采集流程
- refactor(observation): unify cross-frame capture coordination and data ownership | 统一跨 frame 采集协调与数据归属
- The feedback dialog gains "Capture this window": it takes one picture of the app window only when you click, adds it as a removable thumbnail, and never sends it on its own. Refs #11965.
- The session tree now carries each conversation's last-activity time, so a list can sort by recency or show a relative time. Refs #12011.
- A sub-agent queued behind another's write claim now says whose claim it waits on, and whether neither declared write_paths (each treated as writing the whole workspace). Refs #12052.
- Nothing to do. The browser extension updates itself once no chat is busy.
- Find in chat. Press ⌘F (Ctrl+F on Windows and Linux) in an open chat to search the whole chat, including older messages that aren't loaded yet.
- Slash commands. Type / in the message box for /model, /reasoning, /new, /rename, /pin, /status, /review, /init, /usage, /plugins, /pet and /feedback, next to /plan, /goal, /loop and /compact.
- [embedding].max_concurrent defaults to 10 instead of 5. The
- Milvus collections must be rebuilt once after upgrading. They gain the
- Response ids are unique within (app_id, project_id), not globally; the
- proxy: honour RTK_PROXY_BIND in llm-proxy start and the agent launchers (4a583f8), closes
- Projects and conversations: organize research in projects and sessions, explore message branches and side chats, fork a session to continue another research path, and follow plans, clarification…
- Models and agents: choose Claude Code, OpenCode, Codex, or CodeBuddy; connect compatible cloud providers, custom endpoints, or supported Claude/Codex subscription logins.
- Skills, specialists, and connectors: install and update research skills and specialists from their marketplaces, create or import your own, delegate work to specialists, and access scientific…
- Use custom HTTP JSON search providers with configurable authentication, result mapping, and fallback engines
- Rust and TypeScript SDK clients can invoke native runtime commands and transfer sessions without turning commands into agent prompts
- Dismiss side panels directly from the TUI, and keep routine system notices on a compact single line
- Email tab: every application has a new Email tab, between Cover letter and Job posting.
- Only real addresses: the recipient's address is only ever one the posting itself contains. When the posting doesn't give one, the tab says so and you type it in; writing the email again keeps it.
- Change it like the documents: edit or regenerate the subject or the message, ask the chat for changes ("shorter", "more formal"), rewrite the whole email with Undo, or go back to an earlier…
- A commit right after the laptop wakes signs instead of failing: a forwarded ssh agent drops while the laptop sleeps and comes back a few seconds after it wakes.
- You can choose who hears and who speaks. Before this release there was one speech provider and
- The self-check stopped naming a provider the device might not be using. Its key and reachability
- The launcher icon is the character's thinking pose, on white — the same character as before, in
- PyPI: bash pip install crawl4ai==0.9.5
- Docker: bash docker pull unclecode/crawl4ai:0.9.5 docker pull unclecode/crawl4ai:latest
- Note: Docker images are being built and will be available shortly. Check the Docker Release workflow for build status.
- ROB-1237 Price Robusta models whose upstream name isn't provider/id
- ROB-1369 Default the auto executor to one conversation at a time
- ROB-1507 - GPT-6.1 Handle multi-choice LLM responses by merging choices
- core: upgrade DataFusion to 55.1 (563058a)
- Add a shared software-update inventory to Command Center and aos console,
- Expand the canonical Oracle event and policy-routing surface so host adapters
- Treat Finder/FSKit mounting as optional on Darwin. Product install no
- feat(llm-connections): label TypeSafe adapter as "TypeSafe (System One)" in picker
- perf(worker): poll queue metrics once per interval across workers
- ci(notifications): stop sending redundant title, branch, and event to Slack
- Files handles big folders. A folder with thousands of files loads in pages with Load more, and Find a file now searches file contents as well as names.
- Arrange your project sidebar. Choose and order the items in the project sidebar. The layout is saved on your account, so it follows you to every browser.
- Send now on queued messages. Each message waiting in the queue has its own Send now, so you can push one ahead without clearing the rest.
- [OPIK-8753] [SDK] Scope the TS SDK's latest-version check to the prompt's project
- [OPIK-8636] [BE] fix: give Claude judges that think more max_tokens by default
- [issue-8903] [SDK] fix: count cache tokens once for Anthropic and Bedrock LangChain usage
- Recent arXiv papers resolve by DOI. OpenAlex adds new arXiv preprints days late, so looking up an arXiv DOI (10.48550/arXiv.<id>) it doesn't have yet failed.
- When a topic sent to the arXiv source is written in arXiv's own query syntax (search_query=all:..., submittedDate:[...]), only the search words are passed on to OpenAlex.
- <!-- 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)"…
- session identity: order-primary resolver v2 on the wire-neutral domain
- Identity pipeline R2 v1: namespace + origin columns + #2847 proof suite
- Verification stack: #2939 order-centric reconcile merged over the #2815 dev line
- Factory projects now support provider-instance sandboxes (MastraFactoryConfig.sandbox) for settings/builds; existing callback form still works but lacks these capabilities.
- @mastra/server workflow HTTP routes no longer allow starting/resuming internal workflows (returns 404); durable agents must be invoked via agent routes (e.g., POST /agents/:agentId/stream).
- AgentController thread/session behavior changes include new suspension “address” claiming APIs and deprecations: SessionRecord.modeId/modelId/state/pending…
- MCP tools are exposed as mcp__<server>__<tool>; update allow/deny rules and hook matchers that named the old form
- Hooks use the Claude Code contract: matchers, JSON on stdin/stdout, exit codes and timeouts; the old hooks config shape is not read
- Permission rules for Bash match per command, so Bash(git *) style rules apply to each command in a pipeline, not the whole line
- The dashboard's version. The sidebar showed v0.1.0 whatever Hive was installed, because
- hive --version prints the installed Hive's version.
- The first conversation mine after upgrading re-mines each conversation file it covers, once.
- A palace refuses to open with a different embedding model than the one that built it.
- EntityRegistry.research() and confirm_research() are removed. They were an opt-in, off-by-default Wikipedia lookup that nothing in the CLI, MCP server, miners or hooks called.
- Add crewai eval command to report where a stopped model broke.
- Use LiteLLM catalog for context windows.
- Fix flow to learn from resumed human feedback and preserve reviewed output types.
- The cosign claims were validated
- The signatures were verified against the specified public key
- fix(proxy): backport #44243 and #44508 to stable/1.105.x for v1.105.1
- Claude Opus 5, Opus 5.5 and Sonnet 5.5: model support through the Anthropic SDK upgrade.
- Telemetry: fallback attributes lk.fallback.label and lk.fallback.index move to adapter request spans; update trace queries for the flattened hierarchy.
- Anthropic: custom SDK HTTP clients must use httpx2, not httpx.
- Updated dependencies [39cddec]
- @ai-sdk/gateway@2.0.167
- 文件组件改用插件自有的 Rust worker 发布 ops_inspect_local_file、ops_grep、ops_glob 与 ops_replace,提供有界 UTF-8/hex 读取、glob 遍历、Rust 正则搜索,以及带预检计数、内容修订校验与原子替换的 ops_replace。 @T-Auto
- 固定信任的文件操作接入 DSH 原生的策略、沙箱与子进程治理:只读会话可以查看但不能替换,workspace-write 把写入限制在工作区,完全权限不额外附加写入范围,失败时不会退化为无隔离执行。 @T-Auto
- 新增 opt-in 的 dsh-ops/ptc 组件,发布 ops_batch(预检、copy/move/remove、保护路径、写前逐项清单、保守续跑与审计)和 ops_result(有界结果引用、分页、字段/相等过滤、投影与释放)。 @T-Auto
- commit: c950b504649f2c64668c618320c352d159370961
- source tag: beta2026-10-14_11-58-09
- regions: cn global
- Run the three FE/docs-reading tests on FE/docs-only PRs (SKY-19509) (4a34a03a3)
- Show only failing tests' output in OSS CI logs
- Resolve the run type with one task_runs read in GET /v1/runs/{id}/timeline (SKY-17567)
- Support database automations
- Improve super block creation and local column insertion
- Improve database checkbox size and calendar interaction
- Added Auto mode, which approves file edits and safe commands without asking and still prompts before irreversible actions.
- Added the /compact command to compact the conversation on demand and free up the context window: /compact on its own summarizes the conversation and ends the turn, and /compact <message>…
- Added syntax highlighting to the /diff file view, with token colors for each color scheme and a background tint on added and removed lines.
- Automatic collection: documents, web pages, images, and videos your Agents create show up in Artifacts automatically, on one timeline across all Agents.
- Preview and trace back: filter by type or Agent, search, preview in place, and jump back to the conversation that produced a file (Thanks @rudycelekli #3732).
- Pin and manage: pin, rename, or remove artifacts; a file card in a conversation takes you straight to it in Artifacts. The Artifacts page is hidden by default and can be turned on in the menu editor.
- perf(installer): overlap downloads, skip brew auto-update, detach macOS warm-up (68bbec7d7) — Vaibhav Upreti
- feat(shell): guide /loops add input (3337542dd) — Anwesh
- Cookie imports: every cookie-based provider now finds signed-in sessions in all supported browsers, not just Chrome, while Keychain prompts stay bounded: none from background or settings…
- Menu bar: graphical progress-bar tokens and a stacked-bars preset (older app versions keep a readable layout), a single provider icon spanning two text rows, short Cursor lane tags, the Cost…
- Claude: spend-only accounts get a real meter, pace and a reset at midnight UTC on the 1st; reset credits, Extra usage or prepaid Credits, and plan dates are independently selectable.
No release notes published.
Oct 13, 2026
- Run BrowserUse on the host model by default and make browser-use telemetry opt-in
- Let Skills layer project and personal libraries, name each skill's directory, and render a skill for a host command
- Add base_url argument to ZaiProvider
- Act on the whole selection: drag it into a folder, delete it, or insert every selected asset at the playhead from the context menu. Dragging several items shows them all under the pointer.
- Keyboard: arrows move the selection and Shift + arrows extend it, ⌘A / Ctrl+A selects everything in view, Escape clears it, ⌘↓ opens the selected folder and ⌘↑ goes back up.
- Folders open on double-click; a single click selects them.
- Paste or drop screenshots into a session from any browser: pressing Ctrl+V or Cmd+V with an image on the clipboard, or dropping files onto a session's terminal, now uploads them to Glimmervoid and…
- Add or remove repositories in a workspace: a workspace session's restart menu now has Edit repositories, which lists its current repositories alongside the ones found under your repo roots.
- Update checks read the latest release from npm: a release-channel check now asks the npm registry for the published version on every install, including a git clone, instead of reading GitHub tags…
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Fix: C# a nameless member-call receiver (run.Sequence.Apply(), Make().Apply()) no longer mis-binds to the caller's own same-named method.
- Fix: C# new X() no longer binds to a method that shares the type's name, same-file or cross-file; a real same-named type still wins
- Feature: C# foreach loop variables are now typed from the collection's element type (arrays, known sequences, dictionary .Keys/.Values), so calls on the loop variable resolve.
- feat(fivetran): add connector setup, schema config, group, and destination operations
- feat(fivetran): add Fivetran integration with saved credential, selectors, and webhook triggers
- feat(empty-states): add tests and changelog empty-state graphics
- Configure the number of concurrent subagents with CodeInterpreterMiddleware(max_subagent_concurrency=16). The default remains 32 per REPL; requests above the cap queue until a slot opens.
- e0d8311 release(langchain-quickjs): 0.3.9
- f957b8e feat(quickjs): make subagent concurrency configurable
- chore(release): stamp v0.1.8 [skip ci] (8d623a1)
- chore(release): v0.1.8 (986ae71)
- chore(deps): bump async-trait to 0.1.92 (e7a107b)
- The golden image of the inputs 27aac91db9e8, built by the golden image workflow from 4d20e947a130755807e2a8784f5e2dace5753970.
- A new PostCompact hook marks the compacted thread, keyed by session_id plus agent_id.
- That thread's next PostToolUse records the observation as usual, consumes the marker, and returns the timeline once as additionalContext.
- SessionStart clears the marker, so the root thread never gets the timeline twice.
- api: answer a chat to the end on the server with POST /chat/background (e8f2b19)
- api: report the pinned Google tools priming could not find (b49c5f6)
- chat: give each chat in the list its last line (d9d2d55)
- feat(snap): run gateway as a user service by default
- chore(test-guest): remove snap reproduction helper
- fix(ci): publish merge queue E2E statuses
No release notes published.
- refactor(a2a-server): implement V1 to V2 settings migration logic
- fix(acp): bridge PromptResponse.usage and emit usage_update notifications
- Changelog for v0.63.0-preview.0
- Face actions work on about 90% of flat faces. Sketches, offset planes and press-pull now
- Joins: a logo-style part with pieces extruded up and down from one sketch now joins
- Sketch: type a dimension right after picking (D, click, type;
- Fixed duplicate command execution after a lost daemon response so mutating and launch commands are no longer retried after any request bytes have been sent.
- @rotemb-wond
- cli: rtk proxy <cmd> is now rtk bypass <cmd> (the old name stays as an alias), --ultra-compact is removed, rtk gain excludes oversized outputs from its headline rate, and the default binary's…
- cli: pipelines in the CLI, hook gate, rtk llm-proxy (5d8102a)
- engine: pipeline engine, plugin and text crates (62e9047)
- AgentTool.formatApprovalDetails receives ApprovalDetailsOptions; compact: true asks for untruncated bodies so a compact approval prompt can count their lines
- Before-model-call hooks now receive the loop context as a second argument, and buildProviderContext receives the summarizer model
- Fixed a Claude compaction or branch summarizer receiving assistant reasoning from a snapcompact archive rendered under another model, which Claude refuses (reasoning_extraction)
- improvement: [All SDKs] regenerate RPC and session event types against Copilot CLI 1.0.97 (new catalog agent plugin/media type fields and updated session event schemas)
- improvement: [Rust] JSON-RPC and session filesystem dispatch handling refinements
- improvement: [C#] minor update to built-in PermissionHandlers
- bc2be45: Adds t.score(evaluation) for recording a score from 0 to 1, named with .label(key) and gated with .gate() or .atLeast().
- 76f6854: eve's agent spans now come from the shared tracing library, and durable turns recover from one checkpoint per turn.
- f5e5183: Connection approval policies now receive toolAnnotations, the hints an MCP server declared for the called tool (such as readOnlyHint and destructiveHint), so a policy can skip approval…
- Add agent-scoped declarative hooks for custom agents
- Remote plugin marketplace and plugin fetches retry once with your signed-in GitHub token after HTTPS auth fails.
- Streamed responses behind TLS-intercepting proxies complete when close_notify is omitted
- caveman off --all left the Caveman Blocks hook in Claude Code's settings, and every caveman setup added another copy.
- An npm .cmd shim that forwards to a native .exe launches again.
- Hook child processes no longer flash a console window.
- 🐛 An export from the window draws on a device of its own on Vulkan, so it no longer loses the GPU part way through
- 🖥️ An event loop that could not run says why, and the process exits 1
- 📤 A phone's encoder that says "try again" is waited for, frame by frame and at the end
- The daily feed is written quietly, and a routine speaks first. On Android, the iPhone and the desktop the conversation the feed is written in no longer appears in the drawer, the chat list or…
- The Android approval card fits the screen, and the ChatGPT sign-in has a way through a proxy.
- The iPhone, in detail. The conversation scrolls under the header and fades into its scrim, as it does under the composer, and on an iPad the agent's replies line up with the composer.
- Refresh the Core plugin in ChatGPT once. One of Core's tool descriptions changed, so ChatGPT needs to load the new version.
- The browser extension updates itself once no chat is busy; there is nothing to do.
- Chats continue after a delivery timeout. When ChatGPT's answer stops on "Message delivery timed out" or "Connection interrupted", the app now presses Stop and sends Continue on the page as it is…
- Performance: Cut time to first model request by 40% in long sessions. (1754b3e)
- Tools: Intercept and filter tool outputs before returning them to the model using after_tool_callback. (fdda17c)
- Workflows: Require human confirmation before executing workflow nodes via require_confirmation. (988fd1a)
- Significantly faster startup and chat loading — Embedding warm-up now overlaps with chat restoration and startup watchdogs, reducing median boot-to-first-token by ~27% and chat-switch latency by…
- Upgraded LiteLLM (1.104.0) and OpenAI SDK (2.54.0) with an automatic startup migration for existing Docker instances, including installed-SDK HTTP test coverage.
- Native LiteLLM routing for core providers — Nebius, Venice chat, Z.AI/Coding, and OpenRouter embeddings now use LiteLLM's built-in integrations, simplifying provider configuration while preserving…
- fix: reject reusing bundle releases published for another lfx line
- fix(deps): dependency upgrades to 1.12.6
- fix: 1.13.0 security backports to 1.12.6
- fix: point k8s probes at the server's /api/ping
- fix: don't crash when removing a parsed file while an image is attached
- fix: open the clicked image when attachments share a file name
- experiments: coerce average repeated-run latency to float on PostgreSQL (c05a170)
- graphql: normalize retrieval summary counts to int (9735fc9)
- herdr plugin update updates plugins installed from GitHub, all at once or by name. Running plugin panes keep working on their old files, and new runs pick up the update without a config reload.
- Windows notifications stay in Notification Center. Clicking one jumps to its pane and brings the terminal window forward.
- An elevated Windows server now only accepts administrator clients by default. To let normal clients from the same account in, set [server] allow_unelevated_clients = true or start the server with…
- Avoid returning the raw vector column during searches for LanceDB and pgvector adapters — reduces memory use and prevents failures.
- Turso errors raised during database open now surface as OperationalError so calling code can handle them consistently.
- Removed LanceDB background pruning to stop unexpected background work that could interfere with searches.
- Add the Italian interface language
- Give the account service build a 6 GB Node heap
- Open invitation links in OpenBot on Windows
- Docs: Export as Markdown from the File menu, with pictures written to a <file>_files folder (requested
- Sheets: a View-tab toggle for right-to-left grids (@tokg-venv); typed text stays in the Find & Replace "Replace with" box (reported
- Slides: the AI keeps an existing box's font size when it rewrites its text; connector endpoints follow rotated and flipped shapes; the blank theme carries per-script East Asian fonts (@tokg-venv).
- A new outline along the left edge of the conversation lists each prompt you sent. Hover a mark to preview the prompt, and click it to jump there
- Plugins that fail to load or error during a session now show a status line in the chat naming the plugin and the error, instead of silently losing their tools
- You can cancel a Cline Cloud sign-in while it is waiting for the browser
- 547decc: Require Python 3.11 or newer (Python 3.10 is no longer supported).
- Updated dependencies [547decc]
- Updated dependencies [775f017]
- Working folder: a Bot can work directly in a folder you choose instead of the team workspace. Set it in the Bot's settings or details panel: type a path, pick one of your DSH projects, or browse.
- A new Bot asks where to start in its first reply instead of on a question card, and follows its own instructions where they say otherwise.
- Bots no longer use DeepSeek Harness's own question card, which held the Bot until you answered it while messages from other Bots waited.
- 插件命令改名:/plugins 改为 /plugin,本版已移除旧命令,没有保留弃用别名。手动命令、快捷调用和自动化请改用 /plugin,例如 /plugins check … 改为 /plugin check …。(
- TUI 设置迁移:/settings 的 dsh-tui 分区改存 ~/.dsh-tui/settings.json,DSH、Claude、Codex 共用。首次启动自动导入 profile 补丁和旧 ~/.dsh/settings.yaml 中的可编辑字段;!!js 表达式不导入。旧 profile 中残留的字段仍可能在 DSH 内核下提供部署默认值,建议迁移后清理重复配置。(
- 启动与恢复更严格:默认从本包入口启动;DSH_TUI_HOST_ENTRY=0 可恢复原启动方式,DSH 宿主模块不可用时自动回退并说明原因。裸 --resume、-c、--continue 的恢复目标必须属于最终所选后端;没有可恢复会话或后端身份不匹配时明确报错并非零退出。安全模式的“重试正常启动”仍保留告警后冷启动的退路。(
- fix(docs): overview hero reaches the window edges on wide screens
- docs: changelog and blog post for v0.10.3
- docs: fix the home page link preview
- Release notes — formatted for people.
- Changelog — plain Markdown for AI agents and tools.
- npm package
- #2571 1bd2bc5 Thanks @mattzcarey! - createAI (agents/models/ai-sdk, agents/models/pi-ai) adds the zdr gateway option, and vendorKey: "request" for AI SDK models routed with your own vendor key.
- #2572 b7c6ab9 Thanks @mattzcarey! - createAI (agents/models/ai-sdk, agents/models/pi-ai) takes AI Gateway dynamic routes as ai("dynamic/<route>"). See docs/agents/models.md.
- #2573 80719a5 Thanks @mattzcarey! - Update @modelcontextprotocol/client and @modelcontextprotocol/server to 2.2.0 and @modelcontextprotocol/sdk to 1.31.0. See Model Context Protocol.
- 3bacc41f848870b165ddf2cdc83ca8ad76377615 MUL-7573 fix(lark): clear the Typing reaction without trusting in-process state
- 3647ec9a9cc7c198b65b5497038fc3dc07ac2825 MUL-7687 fix(wecom): push an issue notification through the acting agent's bot
- 3d021489f4cacc362f411fb46dd7dc32153d2ca9 MUL-7697: fix(daemon): keep host TMP/TEMP for every Windows task
- feat(auth): enable ChatGPT plan reviews through public Responses
- feat(acp): support ACP service for OpenCodeReview
- fix: follow up ChatGPT auth recovery and provider credentials
- feynman search set works with every search provider pi-web-access supports. It accepted only auto, exa, perplexity and gemini.
- pi-web-access 0.40.0. Asking about a YouTube or local video works again when a thumbnail comes back with the transcript, excerpts no longer cut an emoji in half, and a new opt-in setting…
- #10(@xiangyingchang):将间接依赖 shell-quote 固定到 1.11.0,修复 GHSA-pqg4-j6r4-53mv;appium-ios-device 保持 3.1.24。
- #11(@sunjinshuai):统一已安装 App 清单校验与解析,支持 apps / installedApps 的列表或字典响应;异常响应正常回退,不错误声称已安装或缓存失败结果。
- #20(@cubxxw):长文本续传重新确认原 App、WDA 会话和输入框焦点;上下文变化或无法确认时停止,不向另一输入框写入剩余文本,batch 同步停止后续步骤。
- We're live on Product Hunt. A small card in the app links to MausBot's launch page; it shows once and goes away for good when you close it.
- Bots ask you to sign in where they work. When a site needs a login, the bot shows a Sign in needed card and waits.
- Saved logins. Save a login from that card for its exact site and the bot signs in with it from then on, without ever seeing the password. Settings lists your saved logins per bot.
- @mastra/playground-ui: old Metrics card views/components were removed in favor of the new self-contained Metrics cards/dashboard building blocks.
- @mastra/playground-ui: MetricsCardGroup no longer accepts variant; MetricsLineChart dropped xAxisInterval/xAxisMinTickGap.
- @mastra/memory Observational Memory defaults changed: OM now defaults to model: 'auto' (instead of google/gemini-2.5-flash), which may select a different provider/model depending on available…
- The overnight pass no longer repeats itself. It runs once a day for each project, then rests.
- A script that starts Claude Code or Codex for one small job is no longer saved as a memory.
- A busy session is saved in one go instead of turn by turn.
- 🔀 Review GitLab merge requests and Azure DevOps pull requests alongside GitHub in the session panel, with checks, comments, diffs, and automatic tracking of requests created or updated by agents…
- 🤖 Install reusable custom agents with omnigent agent add, manage them with list and remove, or import a bundle from Create custom agent.
- 🛠️ Browse each harness's MCP servers, skills, plugins, and startup settings in Settings → Harnesses (#9408, @hzub)
- [OPIK-8747] [INFRA] fix: publish the model registry YAML to the CDN only from main
- [OPIK-8721] [BE] fix: put MCP server instructions and deferred tool listings on their server's breakdown row
- [OPIK-8329] [BE] feat: run saved Ollie chart widget queries by reference
- Adds BatchWriteOptions.skip_conflicts (required by VikingBot Compile finalize) and grep(level_limit).
- 💻 Windows:双击下载的安装包,按提示完成安装。
- 🍎 Mac:打开下载的安装包,将 toonflow.app 拖入“应用程序”,再从“应用程序”打开。
- 📱 Android:在手机上打开 APK,按系统提示允许此来源安装应用。后续版本可直接覆盖安装。
- feat(web): require Turnstile on password-reset emails
- feat(ci): bump dependabot to opus5.5
- feat(otel): extract media before TypeScript parsing
- Recovers from runtime trouble on its own. Startup repairs backend, settings and network failures without a reinstall; runtime failures are isolated so the app keeps responding.
- Large session and notebook snapshots transfer reliably, so restoring or moving heavy workspaces no longer stalls or fails partway.
- The Linux command line now also ships as standalone .deb and .rpm packages alongside the existing archives.
- Added a --json flag to agy plugin list for machine-readable output, and fixed the command showing only imported plugins: it now lists every installed plugin, including ones installed from a…
- Changed --model and /model to accept a base model name without --effort: instead of failing with an error asking for --effort, the CLI now picks the model's default thinking level…
- Improved Remote Control so the browser's Settings modal opens for CLI sessions and shows the settings the CLI actually uses…
- feat: web UI notice when a known host-side metering plugin (dsh-context) reports pre-compression context
- release v0.1.195
- fix(ci): explicit concurrency=4 + parallel static-gates — CI total 16min→11m43s (monitored)
- CLI archives built from the Studio release studio-v2.38.0.
- feat(shell): guide goal set input with editable condition form (7f23cada2) — Enoch
- fix: authenticate installer canary metadata requests (72dd884bb) — Anwesh
- fix(security): patch Dependabot alerts and clear open CodeQL findings (3e2c99590) — Vaibhav Upreti
Oct 12, 2026
- skills: reference material for the reflection LM. Pass a path, a directory containing SKILL.md, an inline string, or a list of these. YAML frontmatter is parsed at construction.
- additional_instructions: extra guidance applied to every proposal.
- base_instructions: replaces the reflection prompt but keeps its input and output fields.
- Per-key OpenRouter model lists: with an OpenRouter key, Pi lists only the models that key may use, with the key's context size, max output, and prices, including regional endpoints set in…
- Published configuration schemas: add $schema to settings.json, models.json, keybindings, and themes for editor validation and completion. See Settings, Models, Keybindings, and Themes.
- Clipboard copies over remote sessions: PI_CLIPBOARD_OSC52=1 sends copied text to a remote client's clipboard when Pi cannot detect the remote session, such as a multiplexer attached from another…
- Shell permissions now apply only to the exact commands authorized.
- Local context prompts are preserved across compaction.
- The demo now uses valid Claude 5.5 model IDs.
- release(langgraph): 1.2.15
- fix(langgraph): give every message written to a messages DeltaChannel an id
- fix(langgraph): save a cached node's writes like a node that ran
- Feature: Java a call on a chained, new or cast receiver (factory.builder().size(3).build(), new Builder().size(1), ((Builder) o).build()) now resolves through the declared return types on the…
- Fix: Java a bare call or field receiver inherited from a supertype in another file now resolves to the nearest declaring supertype (class wins over an interface default.
- Fix: C/C++ a quoted cross-directory include (#include "../lib/foo.h", STM32 Core/Src to Core/Inc) now resolves to the right header by a unique longest-shared-directory-prefix match, fail-closed on…
- Make setup a uv tool install instead of a clone
- Make Hermes plugin telemetry opt-in
- Ask before every Hermes phone script and scrub the child env
- feat(web): hint self-hosted v3 deployments to upgrade to Langfuse v4
- fix(web): tidy the v3 version menu upgrade hint
- Policy envelope (security 3.0.5): explicit envelope ceilings require metering, and unmetered worker ceilings stay opt-in (#3503, @rudycelekli).
- Codex MCP backend (codex 3.0.5 and the ruflo wrapper bridge): the opt-in codex backend no longer launches @openai/codex mcp-server, which current Codex CLI dropped, and is retired without a…
- CLI parser, retry, rate-limit, monitoring and config fixes: a reviewed batch (#4107, covering #4000-#4017, @rudycelekli and others).
- npm packages: @on-par/factory-cli@0.9.4 (and lockstep workspace packages) on the latest dist-tag.
- Tag: v0.9.4 → merge commit 4955de02 (PR #2578).
- improve span ingestion reliability (678740c)
- trace: accept empty dataframes in TraceDataset (d0dc9e5)
- Add Scalekit to the Phoenix platform docs (1afe1ea)
- Workspace restore locks change protocol; stop every old process before upgrading.
- Stored tool grants are enforced as written. A grant whose scope.allow list does not match a tool, connection, or application no longer falls through to a broader policy.
- Assistants join your team over MCP — A new "Assistant Connection (MCP)" card in Connections lets Claude Code, Codex, OpenCode, and generic MCP clients connect to Paperclip through browser OAuth or…
- <private> content no longer leaves the hook in server runtime. In server runtime, file edits, tool inputs and outputs, and session-start prompts were sent to the server with <private> sections…
- Older and newer workers can no longer race on a database upgrade. The check that stops an older claude-mem from opening a database a newer version upgraded now reads and records the schema writer…
- A cancelled Chroma connection can no longer release a newer connection's lock. Each Chroma connection now owns its writer-lock token and releases only that lock, so a reconnect during shutdown…
- Constructive sketch solving and live drag: edits keep their solution; value edits that would fold a sketch are refused.
- 3D mice on Linux and six-axis devices.
- OpenSCAD navigation preset.
- feat(pdf): render searchable PDFs #1435
- feat(pdf): preserve hyperlinks and verify native rendering in CI #1437
- fix(paint): restore Muse pill corners and SVG colors #1438
- alphaXiv sign-in messages name the right command. When you weren't signed in, feynman alpha search, get and the paper tools said to run alpha login, a command that isn't installed with Feynman.
- Security: Alpha Hub now requires @modelcontextprotocol/sdk 1.32.1 or later, so a fresh install can't resolve a version affected by GHSA-6qxp-vccf-f47h.
- feat(changelog): publish the Nebius model update
- feat(library): add Stripe billing and failed-payment agent guide
- feat(library): add Stripe payment automation and alert guide
- feat(telemetry): propagate OTel baggage context in Python (900ad33de)
- feat(mcp/ts): add SEP-2663 task execution on mcp/ext-tasks (2c3cfe77b)
- feat(bedrock-ts): support strict tool use via strictTools config (17a33912e)
- fix(docker): copy MARS web bridge modules into the image
- fix(themes): make warning feedback theme-aware
- docs(verify-openhands): map maintenance 2026-10-12
- Changes since langchain==1.4.4
- release(langchain): 1.4.5 feat(langchain): accept summarizer in SummarizationMiddleware
- Removed estimateImageContentTokens() from @oh-my-pi/pi-agent-core/image-tokens; use Tokenizer.imageTokens(size, detail), which prices with the reading model's catalog image rule
- applyShakeRegion() and applyShakeRegions() take the replacement's token count (replacementTokens).
- Fixed Claude context counts reading Opus 5.5 requests about 55% over their billed input: a replayed thinking turn now counts at the output size Claude billed for it, not its signature's bytes
- Added prebuilt Docker images for Linux amd64 and arm64.
- Improved Local uploads, namespace operations, search scoping and CLI commands.
- Added API v5 support. See the API v5 upgrade guide.
- Interactive sandbox settings suggest possible environment secrets and let you add masking hosts before saving.
- Make model IDs in /model and /config model case-insensitive and save canonical IDs
- Keep /allow-all available during startup while enterprise policy is still resolving
- Publish to the community gallery. Make a figure as usual, then say "publish this figure to the Anatomy community gallery".
- Time and tokens, recorded for you. In Claude Code, publishing reads how long the figure took and how many tokens it used from the session's own log, from your request to the request to publish…
- More agents. Anatomy runs in OpenCode and Codex as well as Claude Code, and the gallery knows Antigravity too.
- fix(engine): force reinstall piece when require fails @MrChaker
- fix(worker): installing pieces no longer depends on the worker reaching the public URL @AhmadTash
- Fixed an issue where no fonts would show for the markdown preview configurations in the settings UI.
- Fixed edit predictions retaining single worktree files
- Fixed the remote projects picker sometimes reserving more space than needed.
- Maintenance release — packaging/pipeline fixes only, no user-facing changes.
- The iPad's chat is one column, and the pill bar is glass again. On an iPad the messages, the composer and the pill bar share one centred column of 760 pt, the web console's width.
- Your devices agree with each other. Signing in on an iPhone brings the account's chats, name and face down at once, instead of after a trip to the background.
- A step another device's Muse asks about is answered on the iPhone. A task sent with @〈device〉 runs on that device's Muse.
- feat(learning): name clicked controls with a privacy-safe label (refs PE-712) (6b593cb)
- feat: evaluate request headers when each request is sent (sync and async builders) (b7f0b96)
- fix(sdk-js): fold app context into the leading system message (refs PE-728) (7e97c05)
- #32 60cb113 Thanks @PazChen! - Settings → About in the desktop app shows the version and its update without going to the menu: it looks for one on opening (unless it looked a few minutes ago)…
- #32 60cb113 Thanks @PazChen! - Desktop: when a background task of the agent's own ends after its turn (a command Claude Code ran in the background, a subagent) and it takes it up by itself, the…
- #32 60cb113 Thanks @PazChen! - Desktop: the agent's own commands in the chat, as in its own: type / at the start of a message for the list Claude Code or Codex gives…
- #1068 9e2a8bf Thanks @okwasniewski! - e2e run --watch (-w) runs the tests, then runs again whenever a test file, a source file beside the tests, or the config is saved, or a key asks for it, with…
- #944 9acaa16 Thanks @rudycelekli! - Reject an expired saved session on later restores even when its state is cached in the worker.
- #933 e22d8a0 Thanks @rudycelekli! - Respect model-call cancellation while waiting for a shared subscription token refresh.
- Intents can carry their tickets: a factory-queue-intent request accepts an optional tickets list (id, objective, criteria, dependsOn, writeScopes).
- The factory verifier sees what changed: the verifier that closes an intent now receives the diff of everything that landed on the integration branch since work on that intent began, so it can…
- Factory decision log: the factory now keeps a local decision log at <glimmervoid home>/factory/<projectId>/decisions/<YYYY-MM-DD>.jsonl, retained for 14 days.
- [NA] [SDK] fix: keep a null Bedrock token count from raising in the wrapper
- [OPIK-8733] [BE] [FE] Performance improvements when fetching traces
- The golden image of the inputs f60c6fbd46aa, built by the golden image workflow from a2584f26195062628263f465b4b3ec62939c60d0.
No release notes published.
- The PageIndex SDK, local or cloud — vectorless, reasoning-based RAG, end to end.
- Much faster indexing — the PageIndex Flash engine gets the tree from layout stats: no LLM involved for the structure generation itself, LLMs only write the node summaries, and tree expansion…
- Flash engine: the local default (mode="standard" keeps the classic LLM pipeline).
- Show no percent while a server prepares an update download
- Start a hosted server before the disk restore when a release is ready
- Remove the cloud server purchase from the mobile app
- fix(pptx): a ctrTitle placeholder inherits from the master title placeholder
- fix(pptx): a shape marked hidden is not drawn
- fix(pptx): master decoration is painted below the layout, not above it
No release notes published.
- Durable External Agent Sessions: ClaudeAgent sessions now resume on any replica. The Claude SDK session id is stored on the Agno session, and with a database that supports transcript storage…
- Background Runs & Cancellation for External Agents: ClaudeAgent and CodexAgent support background=True, durable-queue run ids, resumable indexed SSE and cancellation through…
- Step Replay for ClaudeAgent: ClaudeAgent.continue_run / acontinue_run take the same continue_from, fork and input arguments as Agent.continue_run, with stream=True and background=True.
- New projects get a ChatGPT Project. When you add a project folder, the app also creates a ChatGPT Project with the same name, and every new chat of that project opens inside it.
- Background chats in a tab group. In Settings → Agents & automation → Browser & history, Background tabs open in can now be A tab group in your browser window: one collapsed "Chat On Steroids"…
- Workers in their own Git copy. With Isolate new workers in Git worktrees on (off by default), each new worker gets its own Git worktree and branch, so workers don't edit your checkout or each…
- Answers shown once, rendered as Markdown. The TUI no longer prints an answer a second time when it finishes, and a stopped answer keeps the text already shown.
- A TUI that keeps working. A failed request, an unrenderable event, or a restarted service no longer closes the TUI; /reconnect starts the service again when needed.
- Settings that stay current. Default model and access changes show up in the TUI, /new follows your defaults, an approval answered in another client is announced, and /goal clear out … sets a new…
- chore(version): bump version to 2.0.10dev
- docs: announce SOP, TeamPipeline and MiniMax support in news
- fix(tts): label audio with the effective response format
- Fixed intermittent network failures when accessing sandboxes, improving network stability.
- Improved the S3lvol cache, making cross-node restore faster.
- S3lvol hot upgrade: faster I/O recovery during component upgrades.
- fix: allow bugfix-release/* carve-out branches in ci.yml version-guard
- release v0.1.195
- fix: refuse dsh native compaction with 422, not auth-class 403
- core: Keep concurrent credential tests from sharing one node type registry (c17fdee)
- core: Resume tool calls for deferred tools that a skill loaded (fd2586e)
- Updates now follow tagged releases. A source install of the official repository updates to the latest published vX.Y.Z release instead of every commit on main.
- Choose the channel in the desktop app. Settings ▸ About ▸ Updates has an Update channel switch (Stable releases / Every commit) for source installs, and the update dialog now shows which channel…
- Reasoning survives resume. Signed thinking replays after hermes chat --resume and on the next gateway turn, and Bedrock Converse reasoning models no longer get rejected on resumed sessions.
- Updates from 2026.10.1: the first upgrade from stable 2026.10.1 now completes cleanly, successful in-place updates no longer end in errors, and Windows, cloned-filesystem, and private-Node update…
- Doctor and update recovery: Doctor no longer stops on residual legacy imports, stale Gateway locks, slow state inspection, memory reindex locks, or allowlists for removed channel accounts, and…
- Provider-native compaction: OpenAI and Anthropic API-key routes now default to provider-native compaction, and compaction recovers when the primary auth profile is cooling down.
- Tools run on both harnesses. A tool is one module that OpenCode and pi both run. Declare a project's own tools under tools: in kortix.yaml, and limit which agents see them with agents.<name>.tools.
- The Kortix tools are listed in kortix.yaml. web_search, image_search, scrape_webpage, memory and show appear as kortix:<name>. Delete a line to remove that tool.
- One-step MCP install. Installing an MCP app goes straight to the server's sign-in, and a connector you authorize with your own account loads its tools. The app and connector pages are redesigned.
- commit: ebf092dc3fb9dfa3ad25b10258979c0bcd396634
- source tag: beta2026-10-12_12-43-52
- regions: cn global
- CLI archives built from the Studio release studio-v2.37.0.
- bash curl -fsSL | bash -s -- --release --version 0.1.2026.10.12
- Start OpenSRE
- bash opensre
- PDF fidelity restored. Structure extraction keeps complete figures with their captions and legends, reconstructs ruled tables that sit next to formulas without rejecting them, and preserves native…
- A quieter, more robust Windows experience. The desktop CLI works from non-ASCII installation paths, PDF translation handles Windows drive paths, and Notebook kernels and environment captures no…
- Desktop installers and the standalone runtime shed unused payloads — duplicate bundled dependencies, unused model-runtime and tokenizer assets, and platform libraries the app never loads —…
- OpenCode Go models, including grok-4.7 and gpt-6-luna, now route over the Responses API
- Hooks expose turn-end token usage, and turn_start/turn_end now fire for one-shot jcode run turns
- /agents supports per-session worker model overrides
No releases match.
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.
- Nov 16, 2026431 releases · 120 agents · this week
- Nov 9, 2026717 releases · 152 agents
- Nov 2, 2026632 releases · 157 agents
- Oct 26, 2026628 releases · 145 agents
- Oct 19, 2026614 releases · 144 agents
- Oct 12, 2026545 releases · 150 agents
- Oct 5, 2026495 releases · 127 agents
- Sep 28, 2026429 releases · 121 agents
- Sep 21, 2026426 releases · 126 agents
- Sep 14, 2026299 releases · 107 agents
- Sep 7, 2026245 releases · 101 agents
- Aug 31, 2026268 releases · 99 agents
- Aug 24, 2026286 releases · 108 agents
- Aug 17, 2026234 releases · 95 agents
- Aug 10, 2026250 releases · 95 agents
- Aug 3, 2026230 releases · 89 agents
- Jul 27, 2026201 releases · 79 agents
- Jul 20, 2026199 releases · 82 agents
- Jul 13, 2026222 releases · 83 agents
- Jul 6, 2026198 releases · 81 agents
- Jun 29, 2026137 releases · 71 agents
- Jun 22, 2026157 releases · 72 agents
- Jun 15, 2026154 releases · 81 agents
- Jun 8, 2026158 releases · 72 agents
- Jun 1, 2026130 releases · 59 agents
- May 25, 2026123 releases · 67 agents