AI agent releases: Oct 5 – Oct 11, 2026
495 stable releases from 127 tracked agents and frameworks in the week of Oct 5, 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 11, 2026
- ruflo tui: the ruflo console in a terminal of its own (ADR-489, #4059, #4075, #4076).
- MCP resources capability on the CLI server, and a three-host integration loop (Claude Code, Codex, OpenRouter) with ADR-490/491.
- doctor flags a better-sqlite3 built against Node 24.19+ headers on Node 24.19-24.21.
- factory feedback files its issue in your own repo: --repo owner/name, then the new filing.repo config key, then the checkout's origin remote. It no longer falls back to on-par/software-factory.
- The feedback issue leaves out your host name and local design and log paths unless you pass --include-local-paths.
- Upstream factory reports have no built-in target repo either. Without an explicit repo they are skipped with reason no-target-repo.
- The golden image of the inputs d9b4cfd820af, built by the golden image workflow from 6a9cc8c11306f45194de0abdb0bc59f67c55fecb.
- Codex: show saved conversation titles in the Cost menu, retain session IDs in tooltips, and respect Hide personal information. Thanks @obrizan!
- xAI: show an optional live remaining credit estimate from the current invoice preview alongside the posted prepaid balance. Thanks @lilbreadxiaomianbao!
- Menu: add a default-on “Highlight menu cards on hover” toggle to preserve card colors when disabled, while keeping keyboard selection visible. Thanks @cynicalight!
- Safety: graphify uninstall --purge now refuses to delete the project (or home, or a filesystem root) when GRAPHIFY_OUT is empty or ., instead of silently removing the working tree
- Fix: Julia now extracts return-typed, qualified (Base.show), operator (Base.:+) and functor (callable-struct) methods, and resolves broadcast (f.(xs)) and pipe (x |> f) calls
- Fix: Rust grouped use declarations (use a::{b, c::d}) now expand into one import binding per name, instead of a single edge to the common prefix
- 0248e69: eve now downloads attachment links itself instead of handing them to the model provider.
- 0248e69: read_file now shows PDFs up to 20 MiB to the model as files, so the agent can reopen an attached PDF after compaction, or read one that a tool or bash wrote to the sandbox.
- b6b353f: A custom channel's fetchFile can throw attachmentError(message) from eve/channels to tell the model why a file didn't arrive, such as a size limit, while the original error stays in…
- feat(library): How Do You Build a HubSpot AI Agent for Lead Enrichment and Follow-Up?
- feat(library): How to Build a Pipedrive CRM AI Agent for Deal Follow-Ups
- feat(projects): move Project membership to the workspace column
- CLI archives built from the Studio release studio-v2.36.0.
- Improved context caching for repeat turns, reducing redundant prompt processing and making subsequent turns faster when the system prompt is unchanged.
- Renamed TypeSafeJudge to NativeJudge and TypeSafeApiError to JudgmentApiError, including their associated option types. JUDGMENT_ROUTES has been removed.
- Added native judge support for OpenAI’s Decisions API and Cloudflare Workers AI decision models, including the openai-decisions and workers-ai-systemone APIs.
- fix(release-notes): trim 0.1.192/0.1.191 summaries under the 400-char cap
- external-summary: consolidate #2646 + #2661 + #2723 + #2761 (fresh budgets, pool knob, concurrent batches, Go session header)
- test(e2e): real-client cross-protocol switch coverage for pi and dsh
- One look on every device. The bottom bar on Android, iPhone and the web's phone layout is a floating frosted-glass pill, as wide as the composer above it (420 at most, centred on a wider screen)…
- A clip of your own as the face. On Android, iPhone and the desktop the pen on the face and the avatar studio offer Your own clip…: the system's video picker, then a preview of the clip playing in…
- Several custom endpoints at once. The desktop and the web console used to hold one OpenAI-compatible endpoint, so a second one could only replace the first.
- Native companion binary release for oh-my-opencode-slim.
- ChatGPT OAuth Fast 模式(#1502):使用 ChatGPT OAuth 时,可在模型的推理强度面板中开启 Fast 模式;选中后模型选择器会显示 Fast 标记,重新打开会话也会保留选择。提速效果因模型而异,可能增加额度消耗。
- 普通 SubAgent 并发上限(#1506):在「设置 → 通用」限制同一会话内普通 SubAgent 的并发数,前台、后台和恢复的子任务共用上限;达到上限时拒绝新启动,已有任务完成后可继续启动。默认不限,留空保存即可恢复不限;设置对运行中的会话生效,无需重启。主 Agent、Agent Teams 和 Workflow 不计入,也不是跨会话的全局配额。
- 语音听写快捷键(#1510):应用窗口在前台时,按一下开始录音、再按一下结束,也支持按住说话、松开结束。macOS 默认右 Option,Windows / Linux 默认右 Ctrl;可在「设置 → 语音输入」开关或自定义。录音进入当前或最近聚焦的聊天输入框,在其他文本框、终端、弹窗和输入法组词期间不会触发;组合键不会误开麦克风。
- PDF: sign PDFs with a digital certificate (.p12/.pfx or the Windows certificate store) and verify existing signatures (@carloscn).
- Sheets: large streamed workbooks past the Full Load budget can now be filtered on the columns you use (reported
- Slides: canvas gestures match PowerPoint: one undo step per drag, Shift+marquee adds to the selection, modifier keys and guides while resizing, Escape cancels, a 3 px snap pull, and rubber-band…
- assign a random gradient to projects created by ingestion (8946f7a)
- cost-tracking: support Mistral AI model pricing (0e33736), closes
- agents: upgrade pydantic-ai for CVE-2026-107286 (3cdb004)
- Translate all interface text in seven languages
- Keep the Claude prompt cache when only memories change
- Translate open notifications on a language change
- feat(fetch): add curl-compatible HTTP failure flags #1331
- fix(fetch): cancel raw downloads for unsupported output #1374
- feat(fetch): allow explicit interaction before page dump
- #22 387d392 Thanks @PazChen! - Exact frames (render, look, Studio's exports): a video's frame on screen is told by how long its frames last, which the host reads with ffprobe.
- #18 10709ae Thanks @PazChen! - Desktop: a new version, once downloaded, is offered in the top bar beside the projects button (Restart to update) instead of in the chat, where a closed chat hid it.
- #8 6a1d9b8 Thanks @lubobill1990! - Studio: Azure Speech (Azure AI Speech) can be connected under Settings → Providers for voice-over with its neural voices.
- The cosign claims were validated
- The signatures were verified against the specified public key
- feat(guardrails): backport the decision model guardrail to stable/1.104.x
- <path> element: any SVG path data, with viewBox, fillRule and Trim Paths (trimStart, trimEnd, trimOffset). A keyframed d animates the outline, and a path can also be a clip path.
- <ellipse> and <polygon> elements for compositions.
- Dashed strokes: dash length, gap and offset, all animatable, so a dashOffset track makes marching ants.
- Settings → About DS Bot: the DS Bot and DSH versions, Check for updates and the update buttons, and links to the source code, the release notes and the feedback page.
- The update card offers a DSH update only when a DS Bot release names a verified DSH version. DS Bot 0.1.1 offered DSH 0.2.1-alpha.1, which was not verified.
- OpenAI Decisions API support with OpenAIDecisionsConfig and shared decision schemas using the Question marker.
- A2A servers can run on multiple replicas, with cacheable agent cards, opt-in input media type validation, and configurable push notification URL validation.
- Anthropic cache diagnostics and explicit API keys for OpenAI voice configs.
- One-click updates on every surface. codeburn update updates the CLI, the menu bar app and the desktop app in one go. --check lists what is out of date, and --json gives a stable result for scripts.
- The desktop app updates itself and its companions. Click Update: it downloads and restarts on its own, and after the restart the CLI and the menu bar follow. Replacing the app by hand does the same.
- The menu bar update pill updates everything through codeburn update, and checks every 6 hours.
- A standalone Node backend. The shared services behind the desktop, web, CLI and Notebook now run in an ordinary Node process — the command line and headless mode no longer need a windowless…
- Connector reach. A new built-in ENCORI connector brings miRNA–target and RNA–RNA interaction evidence, ChEMBL adds assay details and bioactivity pagination, and ClinVar adds individual submission…
- Fidelity for scientific PDFs. PDF translation preserves native reading order, nested ActualText labels verify correctly, and structure extraction keeps source-owned figures and complete table records.
- [Search] Fix searches for programming languages with hash symbols (e.g., C#, F#) not matching messages.
- [Backup] Fix restore failing on Windows when browser storage is locked by the current session.
- [Errors] Stop misreporting client-refusal errors as "invalid API key" — now shows the correct permission guidance instead of sending users to regenerate working keys.
- Goal on every plan. On plans without a model picker (for example the Free plan), every Goal or Loop follow-up used to fail with "Requested model or reasoning could not be confirmed".
- A background window that stays. With Background chats on, the extension kept forgetting its background window whenever it updated or reloaded.
- Clearer when a tool is on another connector. When ChatGPT runs a script that calls a Chat On Steroids tool its connector doesn't have, the result now names the connector that has it, instead of…
- 66eb494ce197e9b7647e7a4c7cd67b0bad100efa feat: add Marketplace to install app, provider and theme integrations
- 5752bf5bff961d90c082bc5f785ed6069f80ff99 feat: add forward auth endpoint for Traefik, Caddy and Nginx
- 7ddd35543575b04a696ec09e6a3df550ed6fe6c9 feat: add public paths to sites and check site rules in forward auth
- ops_bash 的工具说明明确其只前台执行,没有 run_in_background 与 job id;服务、监听器、watch 等常驻进程应改用宿主 pwsh 工具的 run_in_background,或在命令内部自行脱离。 @Arutoria2024
- ops_bash 现在在 spawn 前拒绝 schema 未声明的参数,错误信息同时点出未知键与可接受的键,避免宿主 bash 与 pwsh 才接受的 run_in_background 被静默忽略后,常驻命令被当前台执行并一直占用到 timeoutMs。 @Arutoria2024
- Windows 系统 PATH 回退现在跳过 %SystemRoot%\System32\bash.exe 这个 WSL 发行版启动器,显式启用回退时会选择真正的 Git for Windows/MSYS bash,避免命令被静默送进 WSL。 @T-Auto
- chat: cut chat-app chats into conversations and search inside chats (3915b6b)
- chat: let a later-registered adapter cut its own old chat history (07d6af5)
- delegation: edit a waiting reply on its card and keep replies out of Drafts by default (c480289)
- Run opencode upgrade --major to upgrade to OpenCode 2.
- Allow the OpenAI ultrafast service tier.
- Validate PTY connection origins and tickets before checking whether a PTY exists.
- feat(repl): share editable input forms across work and cron (d4b9811b9) — Anwesh
- fix(repl): reserve cron list stdin and detach continuing cron ticks (3c3f47765) — Anwesh
Oct 10, 2026
- Release notes — formatted for people, with expandable sections.
- Changelog — plain Markdown for AI agents and tools.
- npm package
- Download the .msi below and run it. The tray app reads everything through the CodeBurn CLI, so install that first
- npm install -g codeburn
- Requires codeburn 0.9.9 or newer and the WebView2 Runtime (preinstalled on Windows 11 and recent Windows 10 updates; installed on demand otherwise).
- Optional OIDC single sign-on, an optional Infisical secret store, and Stripe seat billing (each off unless configured).
- Ordered backup models when a provider fails, and routines that can run on their own model.
- Optional shared ticket board (TICKET_BOARD_ENABLED, off by default).
- fix(gui-app): a failed workspace check no longer reads as "Checking" on the chat Send button (f0dfe6d)
- fix(protocol,cli): agent.list@9.2 carries archived; agent list gains --live and --compact (cfe0c7b)
- fix(gui-app): park an unshown browser guest where the compositor draws it (fdf4b0c)
- Fix: Julia functions with a where clause (long and short form) and parametric type definitions (struct Box{T}, abstract type Shape{T}, struct Sq{T} <: Shape{T}) are now extracted.
- Fix: Fortran a named generic interface now gets a callable node and a dispatches_to edge to each listed module procedure, so a call to the generic name resolves instead of dangling.
- Fix: Elixir defdelegate is now extracted as a function definition, so the delegated function appears and same-module calls to it resolve.
- Complete the Brazilian Portuguese catalog
- Add the Polish interface language
- Do not show the startup error box during a quit or on a server
- feat(changelog): workspace changelog of releases with markdown bodies and the Sim changelog tool
- feat(library): AI Agents for Asana and monday.com Task Automation: Platform and Workflow Guide
- feat(sim-cli): publish which CLI commands Mothership is refused, from the route admission rule
- 592f773: Fix eve build failing with "Identifier has already been declared" when a tool callback's destructured parameter or local declaration shadows a variable from an enclosing scope.
- fix(companion): never store a raw session ID as the agent name
- fix(companion): ship v0.1.4 manifest and guard release sync
- fix(companion): ignore a trailing busy for a finished background job
- test: prevent torn fake Claude captures from blocking releases
- fix(drivers): preserve context notes in chat transcript replay
- feat(sidebar): mark all bot conversations as read
- f34e5253e1829feb8c929896b8eb30bf1fbac6c8 feat(helm): support service annotations
- dcd56517ed05745d6154af02938a5a2589ea551c feat: add expire time to invitations
- 258c80f1d5fbc22962773a37080a5c32adaa42b2 feat: add option to disable self signup but keep provider signup
- Password hash: DOSSIER_PASSWORD takes an argon2id hash of your password instead of the password itself.
- Forgot your password? The sign-in page now says what to do: make a hash of a new one with hash-password, set it and restart.
- Updated state field rendering to trim trailing whitespace and normalize empty collection formatting for consistent prompt output
- Added capability-driven extension terminal launches for tmux, Zellij, Herdr, and CMUX, with consolidated multiplexer detection, provider/shell capability feedback, and required POSIX-shell…
- Added /fork pane|window|tab to open a fork of the session in a new multiplexer pane or window while this session keeps running; bare /fork still forks in place
- Reuse a saved summary. If Compact & Resume stops after ChatGPT wrote its summary (you cancelled it, or the new chat couldn't open), its row in the timeline offers Open a new chat with this summary.
- See why a message waits. A queued message now says when it is waiting for the browser extension to connect, or for a tool call in that chat to finish, instead of showing only a clock.
- The background window follows you. With Background chats on, selecting a chat in the app makes its tab the one shown in the background window. Chrome doesn't come to the front.
- f486188: Windows: an ffmpeg that is stopped (a cancelled render or export, a mix that runs too long) ends with the real ffmpeg
- A render or look frame of a video clip woken ahead of its cut waits for the seek it was woken with before its own
- A team of long-lived Bots. Each Bot is a DSH Session with its own name, role label, instructions, model and character.
- Main Bots that create, change and remove other Bots. You can have more than one Main Bot, change the Main Bot in one step on Settings → Bots, and move one to the front; there is always at least one.
- Messages between Bots, questions to you that do not stop the other Bots, and a view of what each Bot is doing.
- The workflow and craft rules in SKILL.md. Every request is sized first (Figure, Hero or Epic) and the estimate is told to you before work starts.
- The kit: the SVG core, round parts, pipes, the geometry audit, the WebGL layer, React components, and the 3D mode (beta).
- Eight references and the headless Chrome checks in scripts/.
- release: carve out v0.1.192 — #2664 only (anthropic/google dsh rebase fallback) on top of v0.1.191
- Merge pull request #2664 from ranxianglei/2026-10-10_dsh-anthropic-google-rebase
- CLI archives built from the Studio release studio-v2.34.0.
- [Remote Access] Upload images and files to remote Agent sessions from your paired mobile device, with resumable transfers up to 1 GiB per file.
- [Channels] Add WeCom and DingTalk channels with QR setup, Agent conversations, streaming replies, notifications, and attachments.
- [MCP] Support MCP SDK v2 with interactive requests, validated forms, cancellation, and structured results across chat and agent runtimes.
- 初始化和预览请求重试期间保持连接中状态,收到第一帧后显示实时画面。
- 明确刷新连接失败仍显示未连接;保留等待解锁、认证暂停与已有画面的断线状态。
- 更新插件版本和 widget 资源 URI,升级后加载新版界面。
- Ctrl+Enter queues slash and ! commands to run after the current turn, and /clear drops anything still queued
- Alt+G opens the project's git commit log, and diff cycling moves to /diff
- Harness API clients can observe and cancel messages sent mid-turn
- ops_bash 的结果新增 error 字段,并在文本输出中显示 [error: ...] 标记,使 bash 启动失败与命令自身的退出码、信号可以区分。 @T-Auto
- 修复随包 bash 载荷在 profile 声明中存在、但实际已被宿主包管理器清理时,ops_bash 把裸 spawn 异常抛给模型的问题;现在返回可读的 ENOENT 诊断,指出缺失的可执行文件与来源并提示恢复 profile 依赖,其它启动失败也会说明 bash could not be started,且不会静默回退到未授权的系统 bash。 @T-Auto
- ops_bash results now carry an error field and a rendered [error: ...] marker, so a bash startup failure can be told apart from the command's own exit code or signal.
- Replay parity across platforms — mobile replay now handles hub settings, until loops,and empty steps the same way web replay does.
- prefer_intent works on mobile too — the mobile executor accepts the same option as the web executor.
- App launches no longer sink a run — activate_app runs off the main loop and won't fail the run, and the label guard that blocked it is gone.
- The Qoder plugin installs the planning skill and seven lifecycle hooks. Its launcher uses isolated Python startup through Bash, and the adapter maps the optional completion gate to Qoder CLI's…
- Claude Code plugin and standalone skill hooks suppress unchanged PreToolUse views after successful prompt injection. A changed visible view refreshes once.
- File-only SessionStart recovery starts its catch-up child with isolated Python startup, preventing PYTHONPATH startup code from loading in that child.
- Providers: Ollama API keys read the documented balance endpoint (monthly allowance, its reset and purchased credits) and browser mode parses credit wallets; Langdock adds Chrome and Safari profiles.
- Menu bar: optionally follow the frontmost Codex, Claude, Cursor or Antigravity app in the merged icon, color icons by provider, flip the pace sign so reserve reads as positive, and scroll…
- Codex accounts: saved accounts show their own cached usage and refresh individually in Settings, the CLI discovers token and managed accounts through read-only endpoints, explicitly selected…
- commit: da77ac38600311185c17a64b33306827f7cc2878
- source tag: beta2026-10-10_13-00-48
- regions: cn global
- OpenAlex lookups that miss now say why. Looking up a name or title where OpenAlex expects an ID (for example openalex_get_work:Attention is all you need or author: Geoffrey Hinton), or an ID that…
- Automations can now set the agent's model and effort for each run. In the automation editor, open Advanced and fill in Extra agent arguments…
- SSH hosts now run a managed Orca server. Orca ships its own runtime and sets it up when you connect, so connections no longer fail because a host has the wrong Node version or no build tools.
- Native chat (experimental): If you turned on Chat UI with chat as your default view, supported agents now open in the structured chat by default.
No release notes published.
- Added a pinned prompt header in no flickering mode while you scroll back through a long conversation: the prompt that produced the response you're reading stays at the top of the transcript, and…
- Added support for MCP servers to mark a tool as always loaded by setting _meta: {"antigravity/alwaysLoad": true} on it in their tools/list response, so the agent gets that tool directly instead of…
- Added support for a root-level agents.json in plugins, so a plugin can declare its subagents the same way it declares skills and rules with skills.json and rules.json.
- feat(ui): unify record column spacing and metadata hierarchy (bcc3d2a3b) — Anwesh
- fix(repl): restore cron and nested CLI subcommands, retain full cron lists (18f29a7e3) — Anwesh
- fix(shell): reserve stdin for table command variants and correct picker width (a2b3b63c8) — Anwesh
- ghcr.io/omnara-ai/omnara-web@sha256:61f74c188ff3ca409d4390cad87c16507aabe81d3a0be8c6c5e710a8be5435c4
- ghcr.io/omnara-ai/omnara-api@sha256:3e0114834d1ee598ce0147b6150a83e9ab98222d9b63acd647ed65b9c72aaa7f
- ghcr.io/omnara-ai/omnara-worker@sha256:a241a8de40285cf36ed35dbfec44345a3c94cb6d77d92c4da9b692a27987cba7
- An OpenRouter Responses rung (openrouter.ai origin) no longer receives include: ["web_search_call.action.sources"], which OpenRouter refuses with 400 "Invalid Responses API request" while relaying…
- A replayed reasoning item loses its output-only status, as input messages already did. OpenAI refused it with 400 "Unknown parameter: 'input[N].status'" and serves the same item without it.
- Includes 0.7.177: a gateway reasoning carrier stays with its own tool turn when a client records a developer or system message inside it.
Oct 9, 2026
- One tiny image per wiki page. The download count of each asset is how many times that page was viewed. This is not a software release: see scripts/sync_article_pixels.py.
- bb desktop 0.46.0
- Data: your bots work with data. A new Data tab in the Computer panel, powered by DuckDB, with one database per bot on your computer.
- A bot can load files, web addresses or databases into tables, run SQL, and show the result as a table or a chart.
- The result fills the panel, with earlier ones under History. Tables scroll smoothly however big they are, with sorting, a searchable Filter and resizable columns.
- Add XPU to OpenCLIP provider device options
- Add deepinfra as an OpenAI-compatible provider
- Record why an evaluation stopped before its verdict
- pi-web-access 0.38.0.
- Web searches with no API key (Parallel) no longer fail with "free-tier rate limit".
- On a ChatGPT subscription, OpenAI web search now uses the newer, much cheaper gpt-6-luna, so searches take far less of your usage.
- #1003 d1c7fc7 Thanks @ilaigold! - fix: allow jsdom 30 as an optional peer
- #1011 b9471dc Thanks @xpeng5278-web! - fix: treat bedrock throttling as a rate limit for page workers
- #1001 7b419a8 Thanks @Yashwanth-Kumar-Kotla! - fix: give symbol-only headings a readable section anchor so openwiki_search refs to them can be opened with openwiki_read
- fix(sarvam): close the streams the STT collects
- fix(aws): keep realtime session active after stream restart
- fix(telemetry): format a loop stall report without reading source files
- AgentDB floor ^3.0.0-alpha.17 -> ^3.0.0-alpha.20 (memory, cli, root, ruflo). alpha.20 swaps @xenova/transformers for @huggingface/transformers, @ruvector/rvf 0.1.9 -> 0.2.3, ruvector -> 0.2.x.
- Search scores shift slightly (same-text cosine ~0.992 vs alpha.17): an entry that scored just under the 0.3 default threshold can now appear just above it, ranked far below the exact-keyword hit.
- sql.js stale <db>.agentdb.lock: AgentDB alpha.20 refuses saves while the lock exists, and a lock left by a crashed writer was never cleared, so every later save failed and the db file was never…
- Updated dependencies [7b9f900]
- @arizeai/phoenix-client@7.16.1
No release notes published.
- test(desktop): move CSP guards from Rust to the desktop JS suite (5142cf6b2a0f44686eeda1ed07905ee3560fd78e)
- fix(desktop): repair the Windows notification shim's false denial on read (99924a5a69200d34f5cc67e9b76748d7ece72705)
- feat: raise hosted community limit to fifty (8f66f7051c68560feb27dbc74b069c851bc018bd)
- Keeps a gateway reasoning carrier with its own tool turn when a client records a developer or system message between the turn's items.
- Two-factor step-up opens full screen, and the MCP connect dialog uses a split layout with the details beside the form.
- Slack: in a channel thread, the agent runs only when someone tags it.
- The /download page shows the original phone screenshots on the mobile card again.
- add ResourceClaim analyzer for missing DeviceClasses (f2dd92e)
- close the custom analyzer gRPC connection after each analysis (d6d75f7)
- correct API doc lookups for the webhook analyzers (ce78560)
- Fixed a TUI crash when asynchronous questions contain multiple lines, preserving line breaks and complete hyperlink destinations.
- Fixed startup failures caused by differences between a running background server's feature settings and CLI defaults. Compatibility checks now apply only to explicit command-line feature overrides.
- feat(react-core)!: start Trajectory capture for every learning object (d38d016)
- fix(runtime): let COPILOTKIT_OPENAI_API choose the OpenAI API, and log the Chat Completions switch (refs PE-706) (9158409)
- fix(runtime): keep slash model ids intact and use chat completions on OpenAI-compatible endpoints (refs PE-706) (d1bdd92)
- Added a code key to the Claude apps gateway's managed.policies[]: the same settings as cli, also applied in Claude Desktop's Code tab; beside desktop, it turns on Claude Desktop's gateway mode
- Added autoCompactWindow to subagent frontmatter and --agents definitions, so a subagent can auto-compact earlier than the main conversation's window
- Added CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL to run every workflow agent on one model while other subagents keep theirs
- Require Python 3.11 or newer in every package; Python 3.10 installs resolve to 2.54.0 and earlier
- Record the tool requests of ExaSearch, YouSearch, LocalStack and other harness capabilities under durable execution
- Bound retained realtime audio like retained images with retain_audio_max_seconds
- Fix: C# a generic and a non-generic type of the same name (Effect vs Effect<T>) now get distinct node ids (arity folded into the id), so they no longer collapse into one node and Effect<T>…
- Fix: C# a call through a member chain (a.b.C()), element access (xs[i].M()), or inside a constructor body now resolves to the right method, fail-closed to a single declared type per hop…
- Fix: C# a bare call whose name matches an enum member (Delegate(t) next to a Delegate case) no longer binds to the enum member.
- Added an experimental Obscura engine (--engine obscura), launched and cleaned up locally like Chrome with bounded startup and CLI/MCP support.
- Added private CA trust for CLI HTTPS requests so read, install, upgrade and doctor work behind TLS inspection.
- Fixed selector-scoped snapshots (snapshot -s) so elements under ignored wrappers are no longer printed twice and shadow root content of a selected custom element is no longer dropped
- The phone app has its own URL scheme. The system now knows nanoMuse as nanomuse:// (and the iPhone's MCP sign-in callback as nanomuse-mcp://) instead of upstream's minis://, so a phone that also…
- A content check saying no is one sentence, everywhere. When the model provider's content check declines the words of a turn (nanoMuse Cloud's upstream, or the vendor behind a key of your own)…
- The desktop, in three places. A chat this computer had not written in yet, filled by another device's conversation, scrolls to its start.
- 8c55ea7: Skills now load through eve__skill({ name }), which every agent with a skill has in place of load_skill, and a skill with deferred: true…
- The load_skill tool and the eve/tools/load_skill export are removed. Delete any agent/tools/load_skill.ts override or approval policy that names it.
- A skill load's load-skill action now carries the skill's name, and its input is the model's { name } instead of { skill }.
- fix(docx): sectPr page size and margin land in CT_SectPr rank order
- fix(docx): a line grid keeps the paragraph's w:line multiplier
- refactor(pptx): extract shared placeholder/inheritance resolution into SlideComposition
- Show the mobile app card for iPhone and Android
- Show waiting agents in the mobile agent list
- Keep the composer at the bottom of a short chat
- agents: refresh the default models (b747bee)
- chat: name connected tools in plain words and collapse repeated chips (1ce8d97)
- delegation: add an In training step to Handle it for me (f581e8c)
- Use native Microsoft Entra broker authentication on macOS when available, with browser fallback.
- copilot config supports sandbox credential injectHosts keys, with key completion in Bash, Zsh, and Fish.
- context now applies to new and resumed ACP sessions instead of silently using the default or previously saved context tier
- 接入方式:网关是独立进程 openviking-gateway(默认端口 1935),位于客户端和模型服务之间。客户端只改两项:base URL 指向网关,模型服务的 API key 换成网关 key。网关支持 Anthropic Messages、OpenAI Chat Completions 和携带完整历史的 OpenAI…
- 记忆注入:会话开头加入用户 profile 以及记忆、skill 目录;每条新用户消息先检索 OpenViking,再把相关记忆附在这条消息末尾。之后的请求会逐字节回放已注入的内容,provider 的 prompt cache 和 Claude thinking 签名因此保持有效(#5625)。
- 保存与压缩:对话保存到 OpenViking session(名为 gateway-…),由 OpenViking 抽取新记忆;网关新建的 session 关闭 Working Memory。对话接近模型上下文窗口(默认 90%)时,由同一模型写一份有长度上限的摘要,替换切点之前的历史(#5625)。
- feat(helm): make cluster-scoped RBAC optional
- fix(supervisor): keep session retries during startup
- fix(e2e): reserve distinct corporate proxy fixture ports
- Fixes the 1.1.0 server wheel: FastPath stubs now ship in the package and installs import cleanly; SDK/CLI/server packages publish at 1.1.1
- Identity-bound filesystem operations: execd executes file APIs as an explicit Linux UID/GID, with FilesWithIdentity/files_with_identity in all five SDKs.
- Server hardening: snapshot creation submits the runtime object before persisting so early reads can't map to Failed, proxied backend failures map to 502/504 instead of a generic 500, and…
- The app-server reports which process holds a stored session, and pushes a change when that process lets go of it, so a client can show a session open elsewhere as read-only
- Rust CLI: Configure proxy and SSL certificate settings with /proxy-setup.
- Rust CLI: The narrator reads each turn's summary aloud and shows a speaking status row that Esc or Ctrl+C stops.
- 29064bc: ConnectionConfig.get_sandbox_url / get_sandbox_direct_url take an optional port (defaults to the envd port).
- Fixed intermittent text clipping in iOS chats and Markdown previews
- #2484 dedbee4 Thanks @ben-reitz! - Add browserTool() to agents/browser/pi. See Persistent browser.
- #2473 6614955 Thanks @mattzcarey! - Make agents/lifecycle stable along with the Scheduler (agents/schedules), State (agents/state), WebSockets (agents/websockets) and MCPClientManager…
- #2540 b65956e Thanks @mattzcarey! - ThinkHarness now keeps its sessions and operations in the shared agents/harness/store. Migrations are performed on first run. See ThinkHarness: Recovery.
- #947 f3aa4c8 Thanks @kblok! - locator.dblclick() is an alias of doubleTap(), as click() is of tap(), so a test ported from Playwright or Cypress keeps its double clicks.
- #869 1b92a6d Thanks @okwasniewski! - trace keeps its modes but now writes e2e's own trace instead of a Playwright zip: a markdown page per test, on every engine, and the terminal names it under…
- #965 b92658b Thanks @okwasniewski! - A custom StepExecutor that declares vision: true receives agent.assert(..., { vision }) as ctx.step.vision instead of UNSUPPORTED_CAPABILITY.
- 新增简短 AirDrop 指引:手机中的图片、视频或其他文件,通过共享菜单传到当前 Mac,再到下载目录找到文件直接使用。
- 同时发布此前已提交的启动流程改进:首次使用先 setup 检查并复用或启动服务,再取得 READY;启动等待支持有界查询,减少往返回合。
- 验证:Python 回归测试、31 项 widget 测试、构建、插件包校验与 MCP 冒烟检查。
- Added routing-session cleanup for OpenAI Responses and Codex while preserving shared provider fallbacks
- Fixed Claude Haiku 5.5 requests silently enabling adaptive thinking when reasoning is off, on native Bedrock (main and helper calls) and the Anthropic API.
- Fixed /session pin being ignored when every stored account is quota-blocked, which routed the next request to a different exhausted account instead of the pinned one.
- PDF: a Form Design tab to author fillable text, checkbox, radio, dropdown, date and signature fields, with multi-select, arrange and clipboard, editing of saved fields, and auto-detection of…
- Docs: Multiple Pages layout under View › Zoom (requested
- Sheets: Full Load is gated on value cells, so formatted blanks no longer deny filtering in large workbooks (reported
- 💻 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。
- 8584922dc75b0911bf5dea2cbaec64af557f7a45: fix: route newer Copilot models through the Responses API (@DaVinci42)
- Ready for GPT-6. The model picker shows GPT-6, GPT-5.6 and GPT-5.5 with Instant, Medium and High, and Goal and Loop use GPT-6 by default.
- A timeline in rounds. Each turn reads in rounds like ChatGPT's own. The model's sentences between steps are kept, each round is titled with ChatGPT's own summary, the live line says what ChatGPT…
- A clearer Sub-agents list. Each worker is named for its job and shows its own task, its state and its model at a glance.
- [BREAKING] OneCLI installs: upgrade the OneCLI gateway to exactly 1.42.0, which closes a credential-injection bypass.
- /update-nanoclaw now installs the newest release, not the tip of main. The channel is set by NANOCLAW_UPDATE_CHANNEL in .env: stable (the default) follows releases, edge follows main as before…
- Iron Proxy installs: rerun the Iron setup once after updating. It rebuilds the proxy on upstream v0.52.0 (expired proxy certificates are regenerated, host rules match case-insensitively), moves…
- 新增 memory-prompt 与 generation-ref 两类存储,覆盖更多记忆数据结构
- 一键入口 ./start-all-mongo.sh 会把 STORE_MODE 持久化写入 .env,后续启动
- 修复索引冲突场景的自愈能力,重复的默认/抽取记录不再导致启动报错
- Video-domain SDKs are independently versioned npm packages. Default video capabilities still install with Hypit; generic host APIs remain under @hypit/hypit.
- Studio resolves computed CaptionTiming references, displays named Instants even when no Track consumes them, and retains the compact unified Timeline header and projection rows.
- Studio observes atomic Source-file replacement and updates watches as the loaded Source closure changes.
- feat(web): design-system CodeSection replaces CodeView
- feat(sessions): Add syntax highlighting to session tool previews
- feat(billing): send spend alerts for ClickHouse-billed organizations
- [NA] [FE] fix: start the playground with a system and a user message, and keep + Message alternating
- [NA] [FE] fix: say why the output is greyed out after editing a playground prompt
- [OPIK-8720] [BE] Serve a prompt's latest version from the index
- core: Block create/update of workflows with deprecated nodes (ff1a24c)
- Google Gemini Node: Show Nano Banana 2.1 and 2 Lite in the image model list (8dfc695)
- New sessions paint their first frame in about 20ms instead of about 135ms, and the screen no longer shifts while connecting
- Opening several sessions in quick succession is fast again, and resuming large sessions is about three times quicker
- New post_tool_feedback hook feeds a script's output back to the model after each tool call, and post_tool hooks now receive the tool input, call id, and touched paths
- feat: vendor billion-context-pi-subagents in-repo under pi-subagents/
- docs: explain pi built-in subagents and how to disable them
- fix: promptPack=lean actually reaches the proxy system prompt (kernel/packs)
- This release focuses on keeping installed apps rock-solid—especially when loading key modules like search and memory. Onward and upward! 🎉
- — Thank you @senamakel!
- CLI archives built from the Studio release studio-v2.33.0.
- commit: 9fb59bb93cd3953f870b37f405631d914928bb34
- source tag: beta2026-10-09_12-24-45
- regions: cn global
- Guided mode (the default) labels each reply mem-search step 1 of 3 (index), step 2 of 3 (context around up to maxDetails chosen IDs), and step 3 of 3 (one detail batch).
- Automatic mode (mode: "auto") runs the same three bounded stages in one call using ranked lexical retrieval.
- Opaque continuations. A cursor is 27 characters (ms_ plus 24 URL-safe characters).
- Godot 4.2+
- Active Ziva account
- feat: progressively disclose agent tools (2082b19b1) — Vaibhav Upreti
- Require OpenSRE app connections for all GitHub agent work (1a4007686) — Vincent Hus
- fix(llm): raise default output budget to 25,000 tokens (ed975f1fb) — Vincent Hus
Oct 8, 2026
- /v1/messages accepts Claude Code's top-level safeguards field (array of objects, carried opaque) instead of failing with a 400 unknown-field error.
- On a rung whose upstream is Anthropic's official origin (api.anthropic.com, including the subscription path), the field is forwarded byte-for-byte and the beta token is injected.
- The provider's delta.safeguard_results from the final message_delta is relayed unchanged in the client's streaming message_delta, and as a top-level key on non-streaming Messages replies.
- 6189eea: Context compaction now calls its model the way a step does: it streams the summary, retries transient provider failures with the same policy, and counts the summary call's tokens and cost…
- 8c7bb64: Provider options, web search, and prompt caching now read one description of the resolved model.
- 30c88ee: Configure 100% trace sampling when eve creates a Vercel project during eve link, eve deploy, or integration setup, so traces appear in Agent Runs.
- Added /plugin, an interactive manager for the official Antigravity plugin marketplace: browse and install plugins that bundle skills, MCP servers, subagents, rules, and hooks, or install them from…
- Added more options for plugin authors: keep the manifest in a .antigravity-plugin/ or .gemini-plugin/ folder, list skills in plugin.json, store data in a per-plugin folder available as…
- Improved plugin installs: the plugin list opens instantly, an interrupted install never leaves a half-installed plugin behind, and installing or removing plugins on Windows is more reliable.
- This one makes setup work for OpenClaw and keeps Claude Code and Hermes tidy.
- OpenClaw memory now installs on its own. Before, the install stopped to ask for a confirmation nobody could give, so sno setup never finished.
- Claude Code keeps one memory hook for each event, even when its settings already held an old copy.
- Changes since langchain==1.4.3
- release(langchain): 1.4.4 fix(langchain): (SummarizationMiddleware) retry summary step on context overflow chore(deps): bump fsspec from 2026.6.0 to 2026.9.0 in /libs/langchain_v1 chore(deps)…
- Native chat (experimental): With "Use updated structured native chat" turned on (Settings → Chat UI), Grok now opens as a structured chat on this computer and on paired Orca servers.
- SSH workspaces: Open files, live agent tabs and paused Claude or Codex resume offers survive an SSH reconnect and a restart, and tabs you closed long ago no longer come back as empty shells.
- Remote servers and the phone: A paired terminal accepts typing again after its host app relaunches, remote browser tabs come back after a network drop, and the phone keeps every terminal after…
- This release is a big step forward in day-to-day usability: clearer files and sessions, a major memory evolution, and real-time voice agents—plus a ton of reliability and performance work across…
- Imports and migrations are significantly hardened: clearer “building vs empty” states, smarter batching/retries, better failure visibility with retry controls, more reliable resuming, and improved…
- This also includes targeted fixes for web research budgets, credential scrubber behavior (including avoiding code redaction), handling HTTP status correctly, and improved orchestration behaviors.
- Unknown ids and value-less options are errors (#3911, @martinvlad). agent status / task status with an unknown id print "not found" and exit 1.
- Sessions (#3489, #3490, @rudycelekli). session_save captures the persisted task, agent and legacy memory stores by default (explicit false still excludes).
- Backups. The default project backup snapshots both memory.db and agentdb-memory.db and honors the configured memory directory. --db and a non-blank CLAUDE_FLOW_DB_PATH still select a single file.
- Fix: Python nested class identities are now qualified by their enclosing class, so class Outer: class Inner no longer collides with a different Outer2: Inner into one node…
- Feature: Python self/cls/super() calls now resolve to a base-class method defined in another file, fail-closed on multiple/opaque/out-of-corpus bases, with super(Cls, obj) forms excluded
- Fix: Python typed-receiver resolution finds a nested function's file by climbing to the real file node, so a self.x.m()/annotated-local call inside a nested def now resolves instead of failing the…
- Fix ownership retention of rwlock when acquisition is interrupted.
- Resolve issue with documentation site source URLs not being preserved.
- Fix task output display in log-tasks-outputs to show actual output.
- Several Google accounts for Antigravity. Add more than one Google account inside the Antigravity card. Each keeps its own profile, and you pick the account per bot in the model picker.
- Accounts live in their provider cards. "Add Claude account" and "Add ChatGPT account" now sit inside the Claude and ChatGPT cards, not at the bottom of Settings.
- Mentions say who they reach. In the @ list, each bot shows its title, and @everyone says how many bots it will reach.
No release notes published.
- feat(apps): add authenticated backend view helper
- feat(chat): render empty-response corrective nudge as an informational note
- feat: verify-openhands skill with control-openhands CLI and the full feature map
- What's new, after every update. The first time the app opens on a new version, a short summary like this one appears once, with a link to the full release notes. A fresh install shows nothing.
- Clear stops for stuck messages. If ChatGPT never picks up a message waiting in the queue, the app reloads the chat at most three times (after 2, 5 and 10 minutes).
- A new View menu. The "View" text in the title bar is now an icon button next to the sidebar toggle.
- Add Claude Haiku 5.5 to model selection and --model completions
- copilot mcp add recovers cleanly after interrupted MCP config initialization
- MCP enable/disable now works before server discovery without starting MCP servers
- Hypit now has a domain-neutral language host and an adaptive video authoring layer.
- Time is explicit and reusable. Timeline construction resolves an absolute end and anchors; semantic projections are explicit author relations; component surfaces consume named Windows and Instants.
- Captions preserve authored structure. Caption documents distinguish display words, semantic units and authored cues, with complete timing bindings and Studio presentation based on the same model.
- Conversation Compaction: Agent(compaction=True) keeps long sessions inside the context window by folding older turns into a summary.
- Codex External Agent: Added CodexAgent, an adapter for OpenAI Codex built on the openai-codex SDK, alongside the Claude Agent SDK, LangGraph, DSPy and Antigravity adapters.
- HyDE Query Transform: Added QueryTransformer as a new retrieval hook and HyDE as its first implementation: Knowledge(vector_db=..., query_transformer=HyDE()) searches with a hypothetical answer…
- Added onFailure: "block" for command and HTTP hooks: a hook that can't start, times out, or exits with an unexpected code blocks the action instead of letting it through
- Added Program Status Protocol (OSC 7501) support: terminals that implement it can show whether Claude Code is working, waiting on you, or done
- Added quoted text to the /copy picker, so a drafted message copies without its > markers
- Added support for warm-cache-aware conversation pruning, keeping pruned history within the model’s prompt-cache lookback window while preserving Anthropic prompt-cache efficiency.
- Added AgentLoopConfig.hasQueuedAsides (also available on Agent) to allow interruptible waits to detect queued asides without consuming them or signaling other tools.
- Fixed Google Gemini and Cloud Code Assist (Antigravity) requests failing when tool schemas contain unsupported JSON Schema keywords or fields that allow multiple types.
- Add tools for creating and listing managed Git worktrees from trusted local projects when the worktrees feature is enabled.
- Pin tasks in the agent Command Center with p and keep them in a shared Pinned group when supported by the server.
- Navigate and copy transcript blocks with /copy, use Ctrl+Insert to copy selections, and tune mouse-wheel scrolling with tui.mouse_scroll_speed.
- Prompt for goose mode during first-time configure
- HTML session export with an interactive viewer
- Adaptive thinking for Claude Sonnet and Haiku 5.5
- Server side improvements
- External editor swallowing empty buffer
- Models behind OAuth couldn't be used if they weren't in Catwalk
- Updated bundled Claude CLI to version 2.1.294
- CI: alert Slack when a release is blocked or fails
- nanoMuse Cloud is a switch, not a wall. Settings › Models on Android, the iPhone and the desktop starts with Use nanoMuse Cloud models, and the web console's nanoMuse Cloud card has the same switch.
- The main chat follows the chat model you pick. On 0.1.41 a chat pick on Settings › Models set the default for new chats only, and the main chat is never new, so a person who added a key saw their…
- The model pickers fold long lists and can be searched, on every client. Each provider group shows eight models…
- 全新桌面视觉系统「素 Porcelain」(#1452):所有页面重做在同一个视觉系统上——浅色主题、工具调用读成时间线、用户消息用气泡、输入框是悬浮卡片;六个主题统一圆角(4/6/8/12/16/20)和字号(11/12/13/14/15/18/22/26)阶梯,字重不超过 semibold;状态色按语义取(运行中 = info、等待 = warning、完成 = success、失败…
- 手机 / 平板 H5 按触屏重做(#1460、#1461):抽屉式布局换成单层导航栈,首页是按状态分组的会话列表(「等你处理」置顶),系统返回手势回上一层;输入框只留一行(+ / 模型 /…
- 输入框「+」菜单重组(#1462):一级分为「给这次任务用」(技能 / 连接器 / 文件)和「运行方式」(Agent Teams、操作电脑两个开关),其余进「更多工具」;桌面端二级从右侧展开;连接器按「已连接 / 需要处理 / 推荐」列出并显示各自 logo,不再单列「插件」;技能列出最近用过、全部技能和两个可一键安装的市场热门。从菜单打开「操作电脑」与设置里走同一套授权确认。
- docs: changelog and blog post for v0.10.2
- fix(coding-agents): hold the deepen lock with a heartbeat, not a 30-minute TTL
- feat(coding-agents): treat memory as a past record, and let teams add their own guidance
- 插件市场组件拆为三项:dsh-ops-shell 与 dsh-ops-file 默认开启,dsh-ops-background 默认关闭;每项各自导出标题、介绍、图标与中英文本地化。 @T-Auto
- 新增可选的 dsh-ops-background 组件,为科研仿真、模型训练及其他长进程提供后台任务托管;启用后额外发布启动、输出、列出与终止四个后台工具,默认工具面为四个文件工具与 ops_bash,不再发布重复的前台 ops_run。 @T-Auto
- 每个组件独立拥有工具与系统提示词段,关闭即撤销对应内容,不写入 AGENTS.md;文件与后台共享一个按引用计数的插件自有 FastCtx 连接,关闭其中一个不关闭另一个。 @T-Auto
- feat(mcp): expose tool client accessor (1556ede43)
- feat: add vended subagent tool (f076964ee)
- feat(mcp/ts): restore legacy task execution on the SDK 2.x client (d9c971a31)
- feat(docker): create zero vuln base images and add fips mode
- feat(topics): rework the summary system prompt and default facets
- feat(web): show ClickHouse organization id in org debug information
- feat(runtime): tell apart Trajectory setup errors at connect (8bd4389)
- CLI archives built from the Studio release studio-v2.32.0.
- [NA] [FE] fix: move the in-app setup snippets off gpt-3.5-turbo before October 23
- [NA] [SDK] fix: accept doubled-quote escapes in OQL string values in the TS SDK
- [DND-483] [HELM] [CI] Drop the last Bitnami references from the Helm chart
- fix(scan): treat "." as the repository root and convert separators first
- fix(readme): use a proxy-friendly DeepWiki badge
- fix(llm): clarify independent timeouts and deadline diagnostics
- Reviewable Notebook execution. Every Notebook run is screened for destructive or risky code before it executes.
- Resumable shared PDF translation. Translate a full PDF with paragraph-level checkpoints, switch between original, translation, and side-by-side compare reading modes, and resume across sessions —…
- Connector reach. A new PDC connector brings cancer proteomics study, biospecimen and file discovery, and the Genomes connector gains population-specific linkage-disequilibrium queries.
- Docker / Hermes Cloud: nousresearch/hermes-agent:stable (or :latest).
- CLI (git installs): hermes update, or re-run the installer one-liner.
- Desktop app and Termux: no change in this release; they update with the next bundled release.
- agent-framework, agent-framework-core, agent-framework-purview: Add tool-result standing guidance, expose rewritten variable-expansion arguments, and support full-buffer Purview policy evaluation…
- agent-framework-openai: Add the maximum reasoning-effort option
- agent-framework-oracle: Add the alpha Oracle native vector-store connector
- Release notes — formatted for people, with expandable sections.
- Changelog — plain Markdown for AI agents and tools.
- npm package
- Ready to merge comes first in My PRs: pull requests you can merge now sit at the top of the Mine list, and the dashboard throws confetti when one of yours merges, whether you pressed Merge or…
- Coder workspaces stay up while sessions run: set coder.appSlug to an existing coder_app slug to keep a Coder workspace from autostopping while Glimmervoid sessions are running.
- Answered threads show the conversation: each answered thread in a review now shows your comment and the author's reply, whether the automatic check thinks it was addressed and why, and a…
- Pi 1.1.0. Claude browser login no longer fails with "localhost refused to connect" when port 53692 is reserved or in use (common with Hyper-V/WSL on Windows); it falls back to a free port.
- generate and job tools: agents can run any model, price it first with estimate, and poll the job until its files are saved to the library.
- Tool names: the media_ prefix is gone. media_grab is now grab (diffusion grab), and the same goes for probe, transcribe, filmstrip, waveform and listen.
- Agent chat: paste files straight into the message box.
- feat(library): GitHub MCP Server for AI Agents: Builders, Security, and Setup
- feat(library): Best Slack MCP Agent Builders: 7 Options Compared
- feat(library): How to Use the GitHub MCP Server With AI Agents: Builders, Security, and Setup
- The cosign claims were validated
- The signatures were verified against the specified public key
- feat(decisions): backport /v1/systemone, OpenAI-format /v1/decisions and the OpenAI Decisions provider to stable/1.104.x for v1.104.2
- Sheets: a PivotTable Fields pane with report filters and every engine aggregation and show-as mode.
- Docs: Word-aligned line boxes and baselines, East Asian and hangul line factors, floating-table spacing and anchoring, over-page floats, separate cell borders and borderless nested tables, display…
- Slides & pipelines: txBox="1" rectangles classify as text boxes; the notes draft survives an AI deck edit; the deck-spec payload is capped before parsing (reported
- <!-- 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)"…
- 19 个 MCP 工具统一改为 pua_*(Phone Use Agent),常用工具显示 Pua ready、Pua setup、Pua launch app 等标题。
- 同步两项 skill、全部 references、服务指令、恢复 / 下一步调用指引、widget 和命令行入口;保留 WebDriverAgent 底层依赖及既有配置、签名、构建和运行数据。
- 渐变动画顶部 / 底部中央保留微弱光带,随现有 8 秒颜色与 4 秒波点动画自然连接并收回;遮罩只在布局变化时生成,复用透明度交叉渐变。
- SSH remotes start again. Since 0.0.38, connecting to an SSH environment failed with SyntaxError: Invalid character
- On SSH environments, the git branch label no longer opens two new SSH logins every 5 seconds. It checks every 30 seconds now, and switching branches from the picker still updates it right away
- Long model names in the composer show in full when the row has room, and only truncate when it runs out
- Keep previous compression decisions stable across a conversation, including decisions to send raw bytes.
- Pass through rate-limit responses with Retry-After; pin accepted raw retries to their conversation lineage.
- Build the proxy and engine with tree-sitter support for TypeScript, JavaScript, Python, Rust, Java, C, and C++ in the six-platform binary matrix. The container remains a pure-Go build.
- docs(agents): assign new PRs to superset-home for a cloud review
- fix(plugins): list Slack app cards with their title and link
- feat(chat): composer shortcuts, Cmd+Enter steer, and image attachments in the transcript
- 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).
- feat: Claude Code plugin marketplace entry — /plugin install billion-context@billion-context
- docs: neutral JSDoc on noteCompressLoopFailure (content scanners read source that agents open)
- feat(pi): capability-aware proxy-style ACP channel for pi-subagents children
- evals: retry RateLimitError in SyncExecutor instead of failing the whole run (56e9a89)
- For the installer path, run CodeWhaleSetup.exe; it installs
- Extract the archive for your machine: codewhale-windows-x64.zip or
- Double-click codewhale.bat (not the raw .exe) to launch
- ghcr.io/omnara-ai/omnara-web@sha256:1da6755ee4348a62b2bb2f59ac29ff19be8bf8acd3bbe36f9ce3d8566ee3a11e
- ghcr.io/omnara-ai/omnara-api@sha256:91831519638a187e1a4731f07cdde1381458e86747c97d7e893fdb1ef87a43c1
- ghcr.io/omnara-ai/omnara-worker@sha256:9b461913eef215c075310f0f6526c99abd109bc1b8d753d8dc2d04fa59182da0
- fix(terminal): stop echoing a typed slash command under its own prompt row (d37a8ac71) — Anwesh
- fix(terminal): exit the shell on the Ctrl+C the user already confirmed (11d5afe66) — Anwesh
- Tell Slack to connect integrations in the app instead of on the server (7a68d2f3a) — Vaibhav Upreti
- Added a prompt navigator to the chat transcript for jumping between your sent prompts.
- Added MCP OAuth sign-in support, including recovery flows for expired or failed authorizations.
- Combined MCP and skill badges in the chat UI.
Oct 7, 2026
- Snapshot and changelog for v1.15.24
- Capture-only Chat token probabilities (experiential#1249; experiential 0.7.168, exp-gateway-native 0.3.139).
- Add webhook triggers to routines
- Set Haiku 5.5 as the default Claude model
- Show a provider update notice before sending
- Program status reporting: terminals and agent dashboards that support OSC 7501 see whether Pi is working, blocked on a dialog or login, done, or failed. See Program status.
- Claude Haiku 5.5: anthropic/claude-haiku-5-5, with adaptive thinking up to xhigh/max effort.
- Adjust default tools with +name/-name: --tools entries like pi -t +codemode,-write change the default selection instead of replacing it. See Tools.
- Recommend Haiku 5.5 for sidecar models.
- caaa7e7 release(deepagents-code): 0.1.83
- 5d73542 chore(deps): raise dependency minimums for all
- Added Claude Haiku 5.5 for Claude Code 2.1.293 and newer
- Fixed Claude tool calls that run longer than 30 seconds, such as WebFetch or MCP tools, showing up as subagents
- Fixed subagents that finish while the app is disconnected staying "working" until one is opened
- Connecting Whop waits long enough for its tools. Apps reached over the internet now get 30 seconds to list their tools instead of 8, which matters for Whop when you manage several businesses.
- The desktop app runs on a newer Electron (43.7), which includes a fix to its preload sandbox.
- Releases survive GitHub having a bad moment: the release's CI check now retries a GitHub server error instead of stopping. A timing-sensitive test no longer fails at random.
- 内核(kernel):Claude 与 Codex 后端(实验性) — 会话不再只能跑在 DSH 上。用 /kernel(或启动页「内核」格、右下角内核区)打开内核选择器,选中即记住并重启进入,在新内核里开新会话,原会话仍留在 /resume;dsh-tui --backend dsh|claude|codex、配置行 backend:、DSH_TUI_BACKEND…
- Claude:由 Claude Agent SDK 驱动本机 claude CLI,项目里的 CLAUDE.md、设置、hooks、MCP 服务器与插件按 CLI 自己的方式加载。SDK 是可选依赖,未安装时在内核选择器里对着置灰的 Claude 行按 Enter 就能一键装上;凭据依次取 /channel 渠道档案、/login 订阅登录、ANTHROPIC_API_KEY…
- Codex:经官方 codex app-server 复用你自己 Codex 的配置、登录、AGENTS.md、skills、hooks 与 MCP,thread 与官方 CLI 共用(codex resume <id> 可换回官方前端)。需要自装 CLI(npm install -g @openai/codex@0.160.1,基线 0.160.1、最低 0.144.0);/login…
- Added OAuthRefreshUnavailableError, a retryable error that keys.getWithCredential and oauth.access reject with when every usable OAuth credential failed to refresh transiently.
- Fixed a single transient OAuth token-refresh failure (network blip, timeout, 5xx) ending a running session, including subagents restricted to an account pool, with a non-retryable "No API key for…
- Added Claude Haiku 5.5 with adaptive thinking (low through max effort), image input, a 1M-token context window, 128K output, and its tiered pricing above 100K input tokens.
- Added the agent.max_idle_retained_threads setting to control how many idle agent threads remain loaded. (#64547; thanks hazicy)
- Improved Google AI error messages and added automatic retries when a model was overloaded.
- Improved plan updates from ACP agents, including panel refreshes after status changes.
- Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)
- Added agentType to the subagentStatusLine payload, so scripts can tell custom subagent types apart
- Added isDeferred to $.tool.register for mods: false lists the tool's schema in the prompt from the start instead of behind tool search
- MCP HTTP transport. Binding off loopback (--host 0.0.0.0, a LAN IP, ::) now exits 1 unless RUFLO_MCP_HTTP_TOKEN / --auth-token-file / --auth-token is set, or you opt out with…
- Hive-mind gating (#3338, #3339, ADR-476). Remote (HTTP/WebSocket) callers now need an operator credential for spawn, consensus propose, broadcast, shutdown, memory set/delete and optimize-memory.
- Policy ledger anchor (#3602, #3886, ADR-475). Deleting the anchor fields in state.json and truncating receipts no longer verifies as valid.
- feat(core): let a frontend tool resume its pending calls on reconnect (a0480fa)
- feat(mcp-apps): consolidate the Vue and Angular hosts onto the shared package (06f61ae)
- feat: AG-UI 1.0 for CopilotKit (7693a04)
- docs 与 evals 仅保留在本地,不进入 Git、源码包或安装目录;必要屏幕使用说明移入技能 references。
- 精简安装目录的开发测试、CI 与基准 / 冒烟脚本,完整保留 widget 源码、样式、构建脚本、配置、锁文件、测试及已构建页面。
- 中英文 README 补充先在自己的 iPhone 安装、签名并启动 WDA Runner 的步骤,以及 Appium / WebDriverAgent 致谢。
- Fix: more node ids and manifest keys stop leaking the local checkout path or OS username, so a clone keeps producing the same graph — a re-clone no longer keeps a previous checkout's absolute…
- Fix: cross-drive scans no longer abort on a syntax-error warning — the remaining unguarded os.path.relpath in the partial-extraction warning path falls back to a portable display path on Windows…
- Fix: an external import stub whose last referencing edge is gone is dropped on incremental rebuild instead of lingering as a zero-degree ghost a fresh build never has
- @mastra/connect@1.0.0: discovered MCP tools no longer require approval by default; use requireApproval (replaces autoApproveTools) to opt in.
- @mastra/connect: API rename integrations → providers; removed legacy surfaces like the connect() alias of tools(), environment() sandbox credential surface, and several resolver helpers/exports.
- Slack channel integration id changed in channels() from slack → slack-channels (Slack tools remain on slack).
- GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs.
- Amazon Bedrock supports multi-agent V2 and Ultra reasoning on compatible models; Bedrock Mantle also accepts AWS GovCloud regions.
- Sign in to MCP servers from an active terminal session with /mcp login <name>.
- feat(cli): add --timeout-ms alias for fetch timeout #1216
- fix(cdp): misc fixes for cdp tab behaviour #957
- fix(cdp): bound renderer navigation future stack usage #1195
- [Desktop] 目录反复重扫时,删除或归档会话不再一直失败 — 在部分项目里(同一个项目文件夹也被 Studio 2.x 打开过时出现),目录没有任何会话变化,却每隔几秒被重新扫描一次,随后任何删除或归档都会失败,提示「该会话正在执行相同操作」。现在锁文件、临时文件、附属文件和子目录的时间戳不再触发重扫,只有会话记录、元数据和事件日志的变化才会;同一目录的重扫会从 1 秒起逐步退避到…
- [Desktop] 归档旧版恢复行时,整条恢复链一次归档 — 归档旧版恢复链里的某一行时,原来只会处理一到四个快照,每点一次都会留下新的副本,同一批要归档几十次。现在会顺着父节点找到整条恢复链,并一次归档。每个快照之间会释放锁,所以归档很长的链时,其他标签仍可正常使用。个别快照归档失败时其余照常归档,结果会报告还剩几个,再执行一次即可补完。源文件不会被删除或改写,已归档的副本仍可从「已归档」恢复。
- [Desktop · CLI] 只有声明为不读图的模型才会拦截图片 — 没有声明读图能力的模型(自建网关,或目录里没有的模型)以前会被当作不能读图:附加图片会报「没有配置图片理解模型」,而且对话里一旦出现图片(包括 view_image、浏览器截图等工具结果里的图片),之后的每条消息,连纯文字也一样失败。现在只有明确声明为不读图的模型(vision =…
- PubMed titles and abstracts keep their italic terms. Text inside inline tags such as <i>BRCA1</i> was dropped, so a title read "patient with germline mutation" instead of "germline BRCA1…
- PubMed authors keep all their affiliations. An author listed with two or more affiliations came back with none. Thanks to @KennyMcSimpson.
- Security: GHSA-6qxp-vccf-f47h (MCP SDK OAuth client). Feynman's own copies of @modelcontextprotocol/sdk move to 1.32.1.
- [OPIK-8206] [BE] perf: bound trace-keyed spans reads to their spans' weeks; #8767 review follow-up
- build(deps): bump opentelmetry.version from 2.31.1 to 2.32.0 in /apps/opik-backend
- [OPIK-8188] [BE] [FE] fix: reject malformed dataset uploads synchronously and surface failed imports
- Temporal search windows now include their boundary dates and rank date matches more accurately.
- Natural-language searches return matching database rows in the response payload and skip 'turn analysis' to avoid unrelated context.
- Stronger ingestion validation rejects malformed node sets and fixes Notion block rendering (no depth caps).
- Settings › Models: four rows on every client. Chat (the model that talks with you), Operating the screen (looks at the screen and acts for you.
- After a key is saved, *Use it for*. One toggle per thing the provider can handle (chat, the screen, pictures, clips), all on, each naming the model it would get.
- *Automatic*, with one order. The screen, pictures and clips rows open with an Automatic entry that says what it gives right now (Currently nanoMuse Cloud · qwen3.8-27b, say).
- 245d4284a5018b7f446b5eac4d3add1fa5547229 .NET: Restore source-only release solution filter [ #8537 ]
- d0ccaa62761ec36c59c139940797978623fb630a .NET: Preserve explicit null session state across roundtrips
- 279d97f75cb2e7eee885dcdc2c741ef7361ee903 .NET: Honor disabled recent search memory
- New experimental harnesses: AiSdkHarness, ThinkHarness, ContainerHarness (which runs Claude Code or Codex in a Container) and OpenCodeHarness, all with the same shape as PiHarness.
- New agents/websearch: a web_search tool over Cloudflare's Web Search API for pi, the AI SDK and TanStack AI.
- Channels moved to agents/experimental/channels: Channels is being rebuilt around conversations and turns, and the old agents/channels entry point has been removed.
- Usage & Spend grows up: long ranges aggregate by week or month with drill-down to days and a single-day hourly view, source filtering and an amount inspector.
- Codex cost accounting: GPT-6.1 Sol long-context pricing starts at the documented 272K input tokens, Priority turns are priced from durable session settings, drifted ledger mirrors are counted…
- Accounts: Langdock joins as provider #91, ClinePass supports labeled API-key accounts, managed Codex accounts can be listed and promoted from the CLI and appear in the dashboard, promotion fails…
- ghcr.io/omnara-ai/omnara-web@sha256:bfdf48e5b2dcea09ed34d5a95e01da764888493d3a1ef751a7e6c86dc10d5cb4
- ghcr.io/omnara-ai/omnara-api@sha256:75199424c87c48f06780d8953abe41a8a803c499fa53fe51f6d5c68e80257c9e
- ghcr.io/omnara-ai/omnara-worker@sha256:5c387c56919f58f56861fdae9f2457a23a3fa70c3a782f1ac288ba9b9741b4b8
- #14891 ce9a04b - Pin worktrees in the Agent Manager sidebar with Shift+click or the context menu.
- #14885 0abb254 - Open chat web links in-app by default when the experimental Integrated Browser is enabled: sidebar chats use a browser tab and Agent Manager uses its browser panel.
- #14890 3453059 - Reduce the height of Agent Manager section headers so more sessions fit in the sidebar.
- chore(engine): evaluate template-expression scripts with the legacy sandbox @MrChaker
- chore(engine): use the legacy sandbox for the shared script session too @MrChaker
- test(desktop): stop a replayed FSEvents edge racing the hysteresis test (ca3b5bd)
- fix(desktop): make the pid.json start-gap test deterministic, log lost watcher edges (78f0b36)
- fix(gui-app): no ghost cursor while a TUI paints a synchronized frame (5226395)
- ROB-357 Remove kube-lineage from Holmes
- ROB-1233 Constrain impact claims in the narrative to observed evidence
- ROB-1107 Bind conversation worker identity to the Conversations row owner
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.1.32 NPM Tag: latest
- The golden image of the inputs b326f2a1d4fe, built by the golden image workflow from 3ce017171c4ebc50e7ba216b47efeb6ea699ae4c.
- feat(topics): classify traces into buckets
- feat: add external media storage integration
- feat(scores): compile experiment-item and prompt-event reads through the query AST
- Glimmervoid overhead in Usage: the Usage tab's Glimmervoid lanes section is now Glimmervoid overhead.
- Your review comments: a pull request's details and its row hover show how many of your comment threads are resolved.
- Answered threads: team review now resolves answered nit threads automatically (once: a nit you reopen on GitHub stays open for you) and judges other answered threads.
- fix: continuity-aware dsh persona anchor — model switches migrate the anchor instead of forking the main lane
- session self-heal: zombie plugin lanes degrade to proxy tools, idle nudges suppress, 524 summaries learn streaming
- fix: stop labeling first-party siblings as third-party compressors in conflict alarms
- If a model's response ends without a recognized finish reason, Cline asks it to continue once instead of treating the response as complete.
- Refreshed the model catalog. The Cline recommended list adds GPT-6.1 Sol, and the free list adds Solar Mini 4 and drops DeepSeek V4.1 Flash and space-bunny-alpha.
- Generate Commit Message now follows your .clinerules, so conventions like language, format, and ticket references apply to commit messages too.
- <!-- 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)"…
- 4ee572a3e31191bb3834b2148bce7145825bf180 feat: add columns parameter to update-role and update-permission
- 00c52f5ddacbc00c38d48fe4cea5f4acfcda532f feat: add container health checks
- bc93b03f9c320ca3eea8674f39f1ea5f4bd6a899 feat: add initData values to the Helm chart
- chore: pin publish workflow toolchain
- filesystem: write_file can create tiny binary images (png/gif/jpg/webp) for upload flows
- fix(agent): redact sensitive data nested below list boundaries in saved history
No release notes published.
- Add GPT-6.1 Sol (gpt-6.1-sol) metadata and refresh the maintained provider
- Add the current Anthropic claude-sonnet-5-5 choice while retaining
- Rebind existing shell-environment values when resuming an eligible idle
- Improved navigation in the /diff file view: ←/→ now open the next or previous file at its first change with the hunk header and leading context in view, each file remembers its cursor and scroll…
- Improved the /diff file view header to show where you are: it now ends with a right-aligned change A/B · file X/Y position, counting the block of added or removed lines under the cursor and the…
- Fixed /rewind in very long conversations listing older messages as (empty message) and failing with an error when you picked one: messages that were cleared to save space now show their original…
- Steer a running turn. A message sent with Enter while the agent works reaches the running turn at its next step, and the turn does not stop.
- Live queue state. The queue above the composer updates from the session stream instead of a once-a-second poll.
- Spend by source. LLM credit usage shows which trigger, member or caller spent it.
- OpenCode workers: A worker started with orca orchestration worker-start --agent opencode no longer leaves its task sitting unsent in OpenCode's input box on a busy machine.
- Sidebar and workspaces: A remote computer you reveal in the sidebar no longer disappears again while Orca saves the setting.
- CSV files: Large CSV files open without loading the whole file at once. You can resize columns, open links, edit cells, add or remove rows and columns, copy and paste ranges, and sort or filter…
- fix(slack): drop the timeout thread reply and wait 30 minutes (288a82456) — Vaibhav Upreti
- fix(terminal-UI): let the transcript grow down before it anchors (ded92820c) — Anwesh
- fix: remediate GitHub code scanning finding #2402 - Fixed alert #2402 in (b7cb861ec) — Vaibhav Upreti
- Godot 4.2+
- Active Ziva account
Oct 6, 2026
- vLLM tier priority (experiential#1214): GatewayRungDispatchPolicy.upstream_priority (default off; adds no identity bytes when unset).
- Updated bundled Claude CLI to version 2.1.292
- CI: security hardening for GitHub Actions workflows that call Claude — egress-firewall runner, network allow list, auto permission mode, and a check workflow to enforce them
- CI: raised PyPI per-file limit guard to 250 MiB to match the updated project limit
- Fix: incremental graphify update and watch no longer re-cluster and re-render on every run after an unchanged rebuild.
- feat: add boxd as a first-class machine pool provider
- fix: wait for Blaxel sandboxes to finish deploying
- @MichielMAnalytics made their first contribution
- #986 dcf2cc5 Thanks @drakeo338! - fix: accept github line anchors on markdown link targets
- #974 d9d7d99 Thanks @xiehuanyi! - fix: strip complete terminal escape sequences
- #929 d316820 Thanks @c020627! - fix: keep every duplicate wiki heading reachable when one already owns the suffix
- feat(mcp/ts): restore legacy task execution on the SDK 2.x client (d9c971a31)
- fix(models): preserve reported cache token usage (c43bdb45c)
- fix(ts): propagate and resume nested agent-as-tool interrupts (356ca4945)
- delegation: let people describe how they write in their own words (593741e)
- onboarding: describe your work first, then ask what's missing (8e112dc)
- providers: default to Claude Sonnet 5.5 and Opus 5.5 (12b89ce)
- fix(cli): display retry progress indicator during connection recovery
- Changelog for v0.61.0-preview.1
- Changelog for v0.61.0
- 安装脚本自动以原名称 iphone_wda 注册标准 MCP,指向安装缓存入口,保留 2 个 skills 和全部 16 个工具。
- batch 提取共享 selector / observe 定义,保留全部 8 类 op / args,低于默认 5KB schema 压缩门槛;运行时闭集校验及手机操作行为不变。
- 新版已安装本机;244 项普通版自动测试和 126 项视觉版回归通过。新启动的真实 Codex 模型枚举全部 16 个普通工具、确认 batch 参数完整,并成功执行原生 metrics 调用。
- Added canonical redirects and JSON and Markdown data formats for agent-readable stats.
- xAI tool results now include supported images, while unsupported image formats are skipped. (@Jaaneek)
- docs(web): add Fledge Alpha Free to Zen docs
- cd4c62b: Publish the changes from the previous release, which was versioned but failed to publish to npm and PyPI.
- fix(mcp): stop session leak and timeout retries
- fix(mcp): bound server discovery concurrency and cleanup
- fix(mcp): forward global variable headers from the environment
- Fixed interrupted runs so assistant message boundaries are emitted reliably, allowing subscribers to persist and recover the interrupted turn.
- Added getOAuthCredentialProvider() to resolve login aliases, such as openai-codex-device, to the provider where their credentials are stored.
- Fixed Ultrafast service-tier billing and usage accounting: GPT-6 Astra now applies its published premium rates—6× on the OpenAI API and 8× included usage on Codex—and is counted toward the…
- Added project/global preset controls, including scoped preset requests and recovery of pending controls when the target session disappears.
- Added native waiting-input attention and live model/variant details, plus improved native window dragging across monitors.
- Updated companion state handling and bundled animation assets.
- Added --marketplace <source> to claude plugin install: adds the marketplace if needed, under the same policy checks as claude plugin marketplace add, then installs the plugin from it
- Added an effort parameter to the Agent tool, so Claude runs a sub-agent at the effort level you ask for
- Added CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS environment variable to set a longer base delay for the backoff when retrying an overloaded (529) request
- feat(desktop): offer Claude effort levels in agent edit and create (9c6fa14aa03358b4188f2ce0a92a9e435a3cb41b)
- fix(desktop): use object-scale-down for inline markdown images (fcf5f044b3af56f3855a417e61e4389fa925c1bd)
- fix(desktop): detect Kimi Code in its default install path (964b556a429c77c43df8429e97e3648f42d47d52)
- The observer conversation is still recycled into a new generation when it reaches its size budget, which comes from the observer context window (CLAUDE_MEM_OBSERVER_CONTEXT_WINDOW) and the last…
- A context-overflow error from a provider still recycles the conversation and continues.
- The output cap (CLAUDE_MEM_OBSERVER_MAX_OUTPUT_TOKENS) and the cleanup after a rejected request are unchanged.
- B turns on the razor: a click on a clip cuts it on the frame under the
- Every sound on the lanes wears a volume line at the clip's level. Drag
- A scrollbar under the lanes shows which part of the cut is in view and
- Bots stop failing between messages. v0.1.96 kept Claude running between turns to start faster.
- Teammates keep going. A finished hand-off from an outside runtime no longer wakes the bot's own engine for nothing.
- A refused action stays refused. When you, or a bot's approval setting, deny a tool call, the turn ends there.
- Operator UI snippet injection is removed. The PAPERCLIP_CLOUD_UI_SNIPPET and PAPERCLIP_CLOUD_UI_SNIPPET_B64 settings no longer inject HTML into static or development pages.
- Keyboard shortcuts are always on, and their settings are gone. The instance-wide General toggle and the per-user Profile preference are both removed, along with the GET/PATCH /api/auth/preferences…
- @-mentioning an agent no longer starts a run. Mentions in task comments are kept as context for the assignee; only explicit assignment and review requests start work.
- bd3150d9a215684e68a8598abda515d16edc89cb .Net: Bump package version to 1.81.0
- a205c428ee03094751dd33943a8e23571b4e410a .Net: Update file handling for FileIOPlugin
- 58ad1a3780159300c766477f258a4cdfa768eedf .Net: Improve Milvus filtering
- examples/file-only-recovery/ provides a halfway checkpoint, completed reference task, receipt template and objective grader.
- dsh-planning-with-files 1.0.1 uses the producer-owned plugin:planning-with-files source for every injection route, retains legacy message deduplication and has a consistent development dependency…
- Manual /plan-goal and topic-handoff guides use the installed resolver, stop on invalid pins or ambiguous selection, and document its legacy-root convention.
- 817334ed69d92e8bcb3c7b24589f1f3b6f5272b9 feat: add a README to the Helm chart
- 69aba14845c38955b741dcade5edf7030050204d feat: let the global admin set the trusted database hosts of an organization
- The agent can now discover marketplace plugins.
- You can open the effort selector during active runs and the cost breakdown from the footer.
- Added --no-tracing and an explicit tracing status indicator.
- release(langgraph): 1.2.14
- Pi 1.0.4, pi-subagents 0.76.1, pi-web-access 0.37.0.
- Pi: syntax highlighting keeps its colors across multiline strings and comments, MCP sign-in works on servers using OpenID Connect registration, and stalled Bedrock streams are retried.
- pi-subagents: when a subagent's result wakes the main session, that turn keeps its normal system prompt (subagent list, MCP servers, other extensions) and the provider's prompt cache.
- iPhone and iPad: the input field stays. Four releases reported a chat with no input field.
- Windows starts again after the update that moved the app. 0.1.39 installed to Programs\nanomuse-desktop where 0.1.30 had been under Programs\nanoMuse\….
- The desktop moves as the phone does. Every light that said "working" — the chat and status dots, the other device's dot, the thinking dots, the microphone, the glow round the screen — breathes at…
- Pin chats. Pin a chat to keep it at the top of its list in the sidebar.
- One menu for every chat and project. Each chat in the sidebar has a single ⋯ menu with everything you can do with it: pin, rename, open in your browser, block or remove it, and now Copy link to…
- Search from anywhere. ⌘K (Ctrl+K on Windows and Linux) opens a search window. Picking a result opens the chat right at the message that matched.
- Auth on by default: secret generated into ~/.agentmemory/secret when AGENTMEMORY_SECRET is unset
- Writes need Content-Type: application/json; a present Origin must be loopback, the viewer, or in VIEWER_ALLOWED_ORIGINS
- Import and export paths confined to the data dir, ~/.agentmemory, or AGENTMEMORY_IMPORT_ROOT
- See the model behind each answer. A session shows which model answered and what the answer cost.
- pi 1.0. The pi harness is on pi 1.0. It brings compaction and slash commands, and pi now runs only inside the sandbox.
- Config releases follow your repository. A release is a checkout of the base branch with the repository's own layout.
- #851 dc2afca Thanks @KrzysztofMoch! - e2e mcp sessions are now headless by default, the same as e2e run and e2e explore, and the agent decides per session: open_session {headed: true} shows the…
- #790 0d1ed15 Thanks @okwasniewski! - Breaking: e2e and every @e2e-dev package now need Node.js 22.22.3 or newer on Node.js 22, or 24.8.0 or newer (engines.node is ^22.22.3 || >=24.8.0).
- #840 b33398f Thanks @Mohit20057! - screen.swipe({ from, to, duration }) adds an optional duration for path swipes, controlling how long the gesture takes in milliseconds.
- feat(ai-gateway): expose resolution context in response headers, spans and logs
- fix(web): zinc neutral tokens, surface tokens, observation line and fill colours
- fix(evals): skip Bedrock URL media interpolation
- Blog: a new post about skills and desktop apps
- feat(server): let the agent name chats and workspaces via set_title
- feat(state): replace session activity with normalized session lifecycle
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.1.31 NPM Tag: latest
- fix: bound web admin session loading (paged list, summary index, lazy detail)
- fix: expose advisory on /__bili/overview so the web banner renders
- feat: add bili antigravity launcher lane
- [OPIK-8566] [BE] feat: free-form SQL post-run check, behind off/audit/enforce
- [OPIK-8556] [BE] fix: report a Python metric whose declared arguments do not resolve
- [NA] [SDK] feat: add an integration for the native Ollama SDK
- The app-server can discover built-in plugins from roots an embedding application ships, alongside the ones inside the package.
- Vibe Code warns when the selected model cannot view attached images.
- Vibe Code sessions list their background processes in a side panel with live status indicators and a Stop action for running ones.
- Full Changelog
- Changed the default Verbosity setting from high to medium, so the conversation view now groups related tool calls and thoughts into concise summaries while keeping commands and responses visible.
- Changed j and k in the /diff file view to move the line cursor like the up and down arrow keys, matching the file list and Vim, instead of jumping to the next or previous file.
- Fixed trackpad and mouse-wheel scrolling in the conversation view jumping back and forth, most noticeably over SSH or inside tmux, by ignoring the sideways part of diagonal swipes and limiting how…
- feat(env): add Environment.close() to release the LMDB environment explicitly, so the same db path can be reopened without waiting for every Python reference to be collected…
- feat(memo): a memo state function can return recollect_after_run=True in MemoStateOutcome to have its state collected again after a re-run, for states that depend on what the function body read
- Rust SDK ergonomics: #[cocoindex::function(memo, batching)] with per-item memoization, the context_key! macro, and connector schemas derived from Rust row types…
- CLI archives built from the Studio release studio-v2.29.0.
- Google login lets you pick which services to connect and adds a built-in calendar tool for viewing, creating, updating, and deleting events
- Guided Google OAuth app setup that detects existing gcloud credentials, validates generated project IDs, and imports the app automatically
- Alt+Up/Down cycles the speed tier between Standard, Fast, and Ultrafast
- Connector reach. GWAS summary statistics discovery joins the human genetics connector, InterProScan gains sequence submission, and the IEDB connector adds receptor evidence searches.
- A better start. The new-conversation screen unifies research starters with the composer, and the session header gains a common-actions menu.
- Requesty provider. Requesty joins the official provider picker with a curated model catalog.
- Threat models are now saved with Standard, Deep, and Diff scan results and generated security policies.
- Scans now check for exposed credentials in source code, including unused code and tests, and distinguish suspected exposures from placeholders.
- Use --model and --effort when generating patches, validating findings, verifying fixes, or matching and comparing saved scans.
- proxy: read HEADROOM_LICENSE and make usage reporting opt-in
- drop the crewai extra to remove chromadb from the lockfile
- compress: add per-message compression diagnostics (fef99cc)
- chore(desktop): bump version to 1.35.0
- fix(host-service,desktop): refresh PR state after a merge from the PR pane, with a pending state
- fix(desktop): add hover background to sidebar archive button
- fix: support gemini-3.8-live and warn on api mismatches
- feat: add sticky llm fallback routing
- chore: remove unused ATTR_START_TIME, ATTR_END_TIME, and ATTR_SESSION_OPTIONS telemetry constants
- 🔐 Sign in to OIDC servers through your system browser from the desktop app, with silent renewal, public PKCE clients and explicit provider endpoints, plus operator-set server names shown next to…
- ⌨️ Keyboard shortcuts for workspace tabs, common views and recent sessions, Ctrl+Shift+L to focus the chat input, a more scannable sidebar agent hierarchy, and optional worktree deletion when…
- 🛡️ Policies gain period-based cost budgets, conversation_id in the PolicyEvent context, and a turn_final indicator for RESPONSE-phase checks (#5211, #8467, #8653, @terrytangyuan, @ArshSiddiqui)
- feat(skills): add fixing-github-merge-conflicts for conflict loops (e30f98943) — Vincent Hus
- fix(cli): open connection browsers from list commands (e3b30e43e) — Anwesh
- feat(cli): open integrations and MCP directly into connection browsers (be0f6bf21) — Anwesh
- feat: outline the back navigation control on settings sub-pages
- feat(settings): bulk-add provider models as LLM profiles
- feat: add direct-prompt Model Router settings
Oct 5, 2026
- feat(ts): port mcp_router vended tool to TypeScript (ec86d614d)
- feat(bidi): normalize token usage events (3973d1cb8)
- feat(multiagent): migrate agent axis policies to SDKs (d87a0718a)
- Security fix for Copilot: When Orca marked a folder as trusted for GitHub Copilot, it could leave ~/.copilot/config.json readable by other users on the same machine.
- Codex on Windows uses your own ~/.codex: Codex in Orca now uses the same folder as Codex outside Orca, as it already did on macOS and Linux.
- Codex workers: worker-start now accepts max and ultra effort for Codex models Orca doesn't list yet, such as the gpt-6.x models, instead of refusing them.
- Added serverToolUses to the result of a mod's turn.step hook: the tool calls the API ran itself (the advisor), each with its id, name, input, start and end
- Added agentId to the tool.check event of plugin hooks, so a hook can tell a subagent's permission check from the main session's
- Added ceiling to the question and verdict a mod's tool.check hook reads, naming the approval an organization requires for a tool
- Tool patterns and --no-mcp: --tools and --exclude-tools accept * patterns, for example --tools read,codemode,'mcp__radius__*' keeps only one MCP server's tools.
- Codemode persists images: tools.read() on an image file now gives back an image block that image() can show. See Call tools.
- Added * patterns to --tools and --exclude-tools, for example --tools read,codemode,'mcp__radius__*'
- Feature: four more language extractors gained type coverage from @rajatnagda45 — Solidity enum values link via case_of instead of contains, PowerShell enum members likewise…
- Feature: Scala class-level context bounds (class C[A: Ordering]) now resolve to references edges, mirroring the method-level handling from 0.9.75
- Fix: Astro files terminate a trailing // comment in one <script> block so it no longer swallows the next block on the same line; byte offsets and line numbers are preserved
- fix(langgraph): fork before replaying an update checkpoint the thread moved past
- fix(langgraph): keep an update_state on an older checkpoint out of its other branches
- fix(langgraph): keep DeltaChannel counters on every update_state path
- Search your chats. A search field above your chats finds any chat by its title or by what was said in it, including chats older than the list shows. Type a few words in any order.
- Rename a chat. Use the pencil on a chat, or double-click its title, to give it your own name in the app. ChatGPT's title stays as it is, and clearing the name brings it back.
- Your Skills from Claude Code, Codex and ~/.agents just work. Skills from your enabled Claude Code plugins, ~/.claude/skills, Codex and ~/.agents/skills are available to ChatGPT, read-only, without…
- GatewayRungDispatchPolicy.priority_overflow_paying / priority_overflow_pro: per-rung priority overflow multiples in [1, 4], read only under an authored bound…
- Let agents suggest a Marketplace app in the chat
- Add features, steps and questions to the openbot.run home page
- Open the section page when you click a header menu item
- feat(webapi): WebMCP support #1038
- feat(playground): add native WebMCP demos #1192
- fix(skills): remove Google from default websearch engines #1164
- No more screenshot crash on off-screen elements — an element centred far outside the visible screen used to fail the screen read with IndexError: list index out of range and end the run on its…
- Mobile sub-steps use the test's planner — a login or setup step handed to a sub-agent now runs on the same planner as its parent test, instead of always falling back to the older one.
- A passed mobile assertion is saved as passed — assertion results no longer get downgraded in the saved run.
- 火宝快捷配置支持国内/国外节点选择:默认跟随界面语言(中文→firemux.cn,其他→firemux.com),一键写入的推荐配置与获取 Key 链接随节点生成
- feat(mcp): support decision-model evaluators
- feat(experiments): annotate directly from comparison grid
- feat(audit-logs): log actor info for audited actions and deletions
- Main first. Only the main conversation syncs by default; side chats stay on the device that made them.
- {device} is working… When another device is answering a prompt, the line says so under it for up to ten minutes (POST /v1/sync/working, a working hub frame).
- No more 413. The relay accepts 16 MiB (was 6) and says both sizes when it refuses; the runtime keeps the newest four screenshots at ≤ 2 Mpx in a request and replaces the older ones with a note.
- Pi 1.0.3. Sign in with ChatGPT no longer fails with refresh_token_invalidated after a request is cancelled during a token refresh, and closing the terminal no longer reports a read EIO /…
- Azure OpenAI is now azure. Pi renamed the provider from azure-openai-responses, and feynman model login now offers the new name.
- [OPIK-8575] [FE] feat: offer max reasoning effort on OpenAI keys set to the Responses API
- [OPIK-8566] [BE] feat: free-form SQL row policies on the local tables behind traces and spans
- [OPIK-6303] [DOCS] docs: document automatic annotation queue population
- Load a skill’s tools only when the skill is read.
- Reset stale encoding when overwriting text.
- dcba425 release(deepagents): 0.7.22
- 💻 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。
- cc29617: Match BuildKit when filtering the template build context with .dockerignore: a leading ** followed by literal characters (e.g.
- 3ba1d05: Retry a 502 response only for operations that are safe to replay: a gateway may answer 502 after the API already processed the request, so a 502 from sandbox creation, fork, snapshot or…
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.1.27 NPM Tag: latest
- The desktop remembers your Cloud sign-in after a restart. Since v0.1.90 it forgot it on every restart on every OS, which showed "unlock your system keychain", hid lending and left a stale device…
- Threads of one bot run in parallel in its project folder. A bot with a project folder could run only one thread at a time there ("another thread is working in this project folder").
- On Auto, bots don't touch a cloud computer. Only a bot that runs on the Computer engine, which needs one, reaches the hosted desktop on Auto.
- fix: bili codex pins embedded mode via probed --no-daemon
- fix: stamp nativeCaller on host-native agent tool forwarding
- docs(configuration): Server Settings parity with src/config.ts surface
- @mastra/playground-ui: anchorTraceId removed from ThreadViewByTrace, TraceThreadPanel, and ThreadTrace; ThreadTrace.LoadMoreSentinel removed.
- Added agent.getMessages() to tool execution context in standard and durable agent loops.
- Added group filtering and generation ordering to observational memory history. For example, getObservationalMemoryHistory(threadId, resourceId, 1, { groupId, sortDirection: "ASC" }) finds the…
- Full Changelog
- fix(mothership): preserve workflow identity in background confirmations
- fix(files): retain known lineage through binary exports and extraction
- fix(code-placeholders): make python bare-placeholder classification and tagged-template edits linear
- API: Let global members grant variable scopes to API keys (cc25c71)
- editor: Show Gateway credits promotions on community nodes (a509d92)
- Reasoning the way each vendor expects it — adaptive thinking, effort levels, thinking budgets, and a real "off" are sent in each model family's native form.
- Three new providers — MiniMax, Mistral, and xAI, for 13 provider types in total. Any other OpenAI-compatible endpoint works through the custom provider, which now also calls Azure OpenAI…
- Tested, not just listed — every built-in provider, plus example configurations for OpenRouter, DeepInfra, Ollama Cloud, self-hosted vLLM, and others, ships with a ctester report in examples/tests/…
- fix(llm): allow unlisted OpenRouter model overrides
- Added announcement cards above the prompt for model launches, deprecations, and other service notices. Cards appear one at a time, newest first.
- Improved the Windows command sandbox so sandboxed commands no longer need administrator rights, and common tools such as Python, Git, and npm work inside the sandbox out of the box.
- Fixed markdown table columns drifting out of alignment when a cell contains emoji such as ⚠️ or 👍🏽 or scripts with combining characters such as Hindi, which affected every session over SSH or…
No release notes published.
- Newest memories come straight from an index. SessionStart used to fetch every observation and summary in the project, including worktrees merged into it, and sort them all to keep the newest 50.
- Project-alias lookups no longer load rows. Every context render resolves which stored project keys belong to the checkout.
- Local first with cloud sync on. The worker builds SessionStart context from the local database straight away.
- fix(models): frontier model check matches the model name only, never the provider route
- feat(reporting): link HTTP exchange evidence
- feat(prompts): require http_exchange_ids for proxy-validated findings
- OpenAI-compatible providers that answer 529 or another 5xx "heavy usage" error are now retried automatically instead of failing the turn, and the error no longer blames your network
- /usage shows real quota windows for Kimi Code, Cursor, and the Z.ai Coding Plan (thanks @ghoker143)
- Pressing Esc while a follow-up prompt is pending stops the current turn and runs the follow-up (thanks @SiavZ)
- fix(scheduler): a quiet tick cannot hide a tool that failed unverified (bb6283041) — Vincent Hus
- fix(scheduler): read cron weekday numbers the crontab way (0 and 7 are Sunday) (e2b9c28c6) — Vincent Hus
- fix(loops): have the CI and code-scanning templates call the repair tools (f658a1644) — Vincent Hus
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