AI agent releases: Nov 9 – Nov 15, 2026
717 stable releases from 152 tracked agents and frameworks in the week of Nov 9, 2026, newest first, with the headline changes from each set of release notes. Our lists keep each project's latest releases, so a very active project's older releases can be missing.
Nov 15, 2026
- Studio Server: Sub-composition head and root tag scans take linear time (ceb969dba, #7866).
- CLI: The engine and CLI share one trusted PATH rule, and a skipped project copy says why (e2430edec, #7801).
- CLI: Cloud render sends one zip when it fits and uploads only the largest media (9bc2b231d, #7859).
- @vercel/agent-runs 1.0 puts its public API under semantic versioning, so it changes incompatibly only in a major release.
- Record an agent.approval span for every wait on a person, including questions, budget prompts, and connection sign-ins, not just tool approvals.
- Decision spans now name their turn and purpose, and record per-question results as gen_ai.decision.results when the trace records outputs.
- An automation is never switched off by failures. After repeated failures it tries less often (hourly, then every four hours) and keeps going, instead of turning itself off.
- Every run that was skipped or held up leaves a row in the automation's history, with the reason.
- A stuck run no longer blocks the next ones. A scheduled run waits at most an hour for a busy bot, and a run has a default limit of 3 hours (12 hours for a room goal).
- fix(desktop): a plain click on a bare pull request link honors the URL links setting
- chore(desktop): bump version to 2.0.8
- feat(desktop): React Compiler on the renderer, plus Vite 8 / plugin-react 6 / electron-vite 6 / Lingui 6.12
- orchestrator: make Approvals the only gate (2525224)
- today: show Home's Done for you rows on What it does (fa58c68)
- workflows: say what a paused run waits on and stop it on a no (d57447f)
- Context meter after a reload or model switch. For some sessions on custom or self-hosted endpoints, the
- Credentials can pair with the wrong URL on Agent builds after hermes-agent #151775 when a providers: entry named like a built-in provider has its own URL and model.base_url differs (#8973, Form 3).
- External memory providers on newer Agent builds. With a Hermes Agent at or after c9de60702c0d (Agent main since Nov 13, canary builds since Nov 14, and the v0.23.2 release candidates), WebUI chats…
- Minimum Python is now 3.11 (was 3.10). Python 3.10 reached end of life in October 2026, and the patched releases of our optional and transitive dependencies no longer ship wheels for it.
- Security: with 3.10 dropped, the committed lockfile resolves Pillow 13.0.0, filelock 4.2.4 and opentelemetry-api 1.45.2 for every supported interpreter, clearing the open Dependabot advisories…
- Test: the backtick-substitution runtime tests now skip on bash older than 4.1 (they rely on BASH_XTRACEFD, added in 4.1) instead of misbehaving on the bash 3.2 that macOS ships.
- Sessions start again when you run the standalone binary through a symlink. After upgrading to 0.61.0, a glimmervoid symlinked into another folder (such as /usr/local/bin) refused every new session…
No release notes published.
- 侧栏收成主线:设置改为「模型 / 能力 / 偏好 / 数据 / 高级」四维分类,定时任务收进能力,多选改成长按菜单,去掉左滑归档。
- 聊天页:顶栏改成三块悬浮玻璃、去掉右上角文件入口;空状态改成居中问候 + 推荐任务胶囊;用户气泡改淡蓝底深字,和回答气泡对称。
- 查看过程:头部收敛成「目标 → 结果横幅 → 折叠技术详情」,步骤里模型原话正确渲染 Markdown。
- Cheaper than vector RAG: for both index and query.
- Faster: trees are created from layout stats.
- Unified SDK: vectorless RAG. Local or cloud, for index and chat.
- App: a start whose Home gestures did nothing while the board's USB is up, but with no mouse on this iPhone,
- App: when no fresh screen image arrives at startup, the app reopens the board's stream once before asking to
- App: the firmware check on Home runs at most once a minute, in a task a reloaded Home doesn't cancel. Some
- Rotate desktop and daemon DSNs (stable hotfix)
- The tree-sitter CLI installs on networks where its download script keeps failing
- A corrupt Chroma collection is dropped once, not once per parallel backfill
- help prints help instead of running the command
- Ships #1365: gateway exits for five stuck states. These are the guardrail pool worker leak on cancel, the alias reload backoff, health registry pruning, closing requests left open by a crash…
- /brag now switches to /brag-slim on Sonnet 5.5 and Fable 5.2, not just Opus 5.5.
- duration works: ask for a length and the video is planned to it. 15–25 seconds is still the default.
- Vertical videos keep text clear of the Reels, TikTok and Shorts buttons and captions.
- prompt_tokens is the whole input, including cache reads and writes;
- cache_read_input_tokens and cache_creation_input_tokens are parts of it, never additions;
- total_tokens >= prompt_tokens + completion_tokens.
- 支持 Helm 集群部署。多实例共用一个 PostgreSQL
- 管理员可以限制创建专家时能选的存储后端
- 新增托管专家,对话可以交给后端 Agent,如 OpenCode
- fix(cdp): generate complete discovery schema at build time
- fix(protocol): let independent commands progress during local waits #2840
- fix(browser): keep omitted window bounds and add getWindowBounds
- Add A7 Director mode with a fixed fine-line character, vintage 2D scenery, reusable prompts, reference stills and production/acceptance rules.
- Include the approved two landscape showcases and two accepted cross-scene workflow examples. Lyrics are showcase-specific, not an A7 default.
- Update all nine README files: A styles on the left, B styles on the right, 13 published styles in total. B7–B10 stay unpublished.
- A run stopped by its time limit always leaves a receipt. When the engine is stuck and the supervisor's backstop has to stop it, a short seal pass now commits and seals the current tree as PARTIAL…
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.3.5 NPM Tag: latest
- Deterministic incremental preparation reconciles file changes, ignore rules, structural fingerprints, and imports.
- Source-evidence checks protect existing symbols during incremental updates, and failed finalization restores the saved graph and fingerprints.
- Fingerprints match methods by their enclosing owner and name, keeping same-named methods in different classes distinct —
- runtime: A full runtime whose orchestrator config names no promptProfileConfig sends no built-in concise preset; pass DEFAULT_PROMPT_PROFILE_CONFIG to keep it.
- runtime: AgentOS.create keeps the runtime's ENABLE_PERSISTENCE value when a host's orchestratorConfig does not set enableConversationalPersistence; before, it was on.
- runtime: The full runtime registers no getCurrentDateTime tool of its own; add createCurrentDateTimeTool() to tools to offer it.
- [PR] gate: fix upstream recorder leak causing real model gate deadlock (pipe inheritance) | gate: 修上游 recorder 泄漏导致真模型门禁卡死(管道继承)
- [PR] chore(branch): land 40 single-purpose fixes as one history-preserving batch (wave 5)
- [PR] fix: map search evidence offsets back to original Unicode text
- 91ba94f10e41229c3f3755fb383a90155b23af19 feat: hide the Casdoor Cloud links and upgrade offers by default in self-hosted community editions
- 72eb9795e01123531348474f42cac50bafe19468 feat: improve plugin logic
- 3d81f45bc98be0fe379b3a21973c783060066af5 feat: improve plugin uninstall logic
- This SDK release bundles CLI version: 0.27.4
- Source: CLI built from source (same branch/ref as SDK)
- This SDK release bundles CLI version: 0.27.3
- feat(review): default --to to HEAD when only --from is given
- feat(cli): render Markdown summaries and style progress output
- feat(review): support --resume for workspace mode
- 语音:朗读、语音输入和语音通话。 回复可以用 DeepSeek 的声音朗读,输入框可以说话输入,也可以直接和她语音通话。通话里能随时打断她,她会自己判断你说完了再接话,有实时字幕;「停顿多久算说完」可以选,默认多等一会儿。通话中可以换声音和识别方式。用账号登录时语音接口不认账号,会改用 API Key。
- 酒馆:玩法包和做卡工具。 新增玩法包;做卡时可以让 AI 改卡、挑着应用改动、看历史版本、看最终提示词,还能先试玩。长对话不再越聊越卡(显示、用量统计、读对话都快了很多)。支持工具回合和自定义模型渠道,数值可以手填;回复长到顶栏下面时顶栏会加遮罩。
- 手机操控:悬浮条和不打扰模式。 操控时屏幕边缘换成柔和的色带;悬浮条可以拖动、贴边,小球闲置时半藏在边上。不打扰模式选好通道、连上就能用,不再被一整套准入检查拦住;主屏读不到当前应用时改成确认一次,不再直接拒绝。
- The models Concat downloads on demand, mirrored from the upstreams named in models/manifest.toml. Not an app release: see the tags beginning with v for those.
- bash curl -fsSL | bash -s -- --release --version 0.1.2026.11.15
- Start OpenSRE
- bash opensre
- feat(library): Microsoft Copilot Studio vs Power Automate: Which Should You Use?
- feat(library): How Do You Build a Governed AI Agent for Microsoft Intune Device Management?
- feat(desktop): quick chat window on a global shortcut
- Restore version 1.0.58 in pyproject.toml and uv.lock, and set it in the SDK client headers (SKY-20765) (f3a86f568)
- Narrow Copilot test-run agent steps to approved accounts by default (SKY-20485) (03c339646)
- Upgrade dnspython to 2.9.2 in the root lockfile (SKY-20641) (cab0de6e9)
Nov 14, 2026
- Visual proof now uploads for factory-opened PRs; the guard refuses only an evidence branch equal to the PR head branch or under factory/.
- Queue preflight now adopts the issue's own factory draft PR left without a draft reason (an interrupted SHIP or a ci-failed park): green or pending CI lands it, red CI queues CI rework, instead of…
- SHIP no longer parks a run just because CI is still pending: the visual-proof gate still classifies (non-UI changes are not required) and captures locally, so SHIP leaves the usual CI-pending draft.
- Engine: Renders log the class of every address the proxy refuses (89bca002e, #7762).
- Producer: Compile downloads composition scripts through the shared asset downloader (cd55f059c, #7763).
- Long task descriptions no longer slow a run down: the check for file paths named in the task took about 7.6 s per call on a 99 KB task and ran several times per run. It now takes a few milliseconds.
- The proof that a test fails without the fix no longer runs the fixed code. On a repo whose virtualenv holds an editable install, the base checkout used to import the already-changed sources, so a…
- Off by default, opt in with LOKI_E10_REVIEW_TURN_GUARD=1: when the pre-delivery review turn edits anything, Loki reruns the tests of what it changed, and takes back only that turn's edits if a…
- Azure: Fixed Kimi K3 on Azure AI Foundry losing its earlier reasoning after tool calls and refusing reasoning_effort
- Azure: Fixed 400 on Foundry project /openai/v1 bases and 404 on resource-level /openai/v1 Responses bases
- AWS Bedrock: Fixed 500 on s3:// image, PDF and video URLs on Converse models, which now go to Bedrock as S3 sources
- Subagents work alongside the main agent and report back when they finish.
- Switch to any subagent to watch it, message it, or answer its prompts.
- A parent agent hears when its subagent's command has printed nothing for 2 minutes.
- Fix: csharp a non-UTF-8 source file (Windows-1252/UTF-16) that contains a preprocessor directive is decoded before the directive is blanked, so a non-ASCII identifier is no longer cut short and…
- Fix: csharp an explicit interface implementation (void IFoo.Bar()) is kept out of overload binding and gets its own node, so it no longer collapses with a public Bar overload, and calls dispatch…
- Fix: ids members or functions of one scope whose names differ only by a leading/trailing _, or only by case, get distinct node ids instead of one swallowing the other, now in Kotlin, Swift, Dart…
- ⚡ Smoother busy workspaces: Crowded sidebars and long chat tool runs render only the visible rows, reducing rendering work and chat memory use.
- 🧹 Clear finished work on your terms: Settings → Workspace Cleanup lets you review matching workspaces, enable automatic cleanup and inspect its activity. Give individual projects their own rules.
- 💬 Follow one subagent at a time: Open its task, live steps and final answer in a dedicated view, then return to the main conversation where you left off.
- Security patch: fast-xml-parser 5.12.1 and fast-xml-builder 1.3.2. Fixes a quadratic-time attribute scan in the XML validator (GHSA-7cfw-w2rf-f3jc) and an uncontrolled recursion crash in the XML…
- Added the Antigravity CLI ACP provider with managed installation, Google sign-in, refreshable distributed credentials, host tools, and retained conversations across tool steps and history changes.
- Added Antigravity CLI models for Gemini, Claude and GPT-OSS with selectable native thinking levels.
- Added the antigravity-cli provider for Google Antigravity over ACP, with managed binary installation, /login antigravity-cli Google sign-in, host tool execution, and retained conversations across…
- The question manager works out of the box. Turning it on for a session now answers a question by picking its one option labeled (Recommended), with no rule to write first and no model call, so it…
- Reviews board columns fit the board. Columns snap into place as you scroll the board sideways, the edge that hides more columns fades so you can tell more sit off to the side, and on a narrow…
- Dragging a Reviews board card opens a slot where it will land. The cards around the pointer slide apart to make room, the card's old place closes up, and on drop the card settles into the slot.
- Windows: SolidCraft-0.42.0-windows-x64.exe runs as is. Or install the .msi, or unzip the
- macOS 11+ (Apple silicon and Intel): open the .dmg and drag SolidCraft.app onto Applications.
- Linux (x86_64 and aarch64, glibc 2.35+): the .AppImage (chmod +x, then run it), the
- Connect Console HITL question parsing, ordered answers and conversation continuation. Display and copy readable answer summaries while preserving protocol payloads for transport and storage.
- Support per-question field validators, control renderers and custom cards. Preserve input for retry after failed submissions and discard obsolete interactions when reloading or switching sessions.
- Unify HITL labels, help text, error tips and invalid-control icons. Clear local errors when input values change.
- Give the start failure reason when the remote desktop runtime has an unexpected error
- Animate the theme change in the mobile app
- Show the tax ID field on the payment page from the start
- 适配 DSH 0.2.2-rc.3 — 主适配线前移到 0.2.2-rc.3,peer 范围同时保留 0.2.2-rc.2、0.2.2-rc.1 等旧验证线,停在 rc.2 的用户也能直接装这版。rc.3 是版本列车重发:共有包的 .d.ts 零变化,唯一 .js 变化是 dsh-tool-cordis 的 API 目录,宿主复刻指纹与 patch surface 内容都没变,本版没有…
- 插件可以往 @ 补全里加行(实验性) — 插件用 ctx.tuiMentions.register({ id, title, complete(query, signal) }) 往输入框的 @ 补全浮窗加行,与工作区文件共用一张卡:出现插件行时卡片标题从「文件」变成「提及」,插件行按与文件同一套模糊分并排(文件顺序不动、平分归文件)。行是 { label, detail?…
- @ 补全的插件行只做文本插入 — 选中插件行只替换当前 token,不参与文件提及展开器;insertText 限纯文本、单行、≤200 字符,含 @ 提及记号的在准入时丢弃。register 对非法输入、重复 id 或超额直接抛错,dispose 随插件 activation 回收。这是实验性接缝,仍可能随 TUI Profile 演进调整;没有插件行时补全浮窗与之前逐帧一致。
- B6 now supports early character preflight, reusable action/camera helpers, lower-resolution 30fps drafts, and changed-input iteration planning through the installed workflow.
- Final native rendering and all existing quality and approval checks remain required. Approved sample videos are unchanged. No test films are included.
- Validated with 223 Python B6 tests, 21 Node B6 tests and an unchanged native render of an existing production. Draft rendering was approximately 3.5× faster in that local benchmark.
- memory: require a retention choice for a souled store more than one person or session shares, and sweep expired traces and their pages
- avatar: take a caller's acceptance check, emotion list and angles in AvatarPipeline (2e5df25)
- hearing: a room's audio output for a host that already reads the room (f5c452e)
- Identify Chat sessions to Claude Code as Orchestrator (stable hotfix)
- Ask a bot to change a setting and it does it, with an Undo in the chat.
- You can make MausBot a regular bot if you don't want it as your helper.
- Point at a member's page from inside a room.
- Every run is judged by your own model: one more low-effort request checks the run against the task and fills result.judgement with success, failure_reason, description, prompt_injection_detected…
- The judge sees the trace, the run's last screenshots and the final page (unless use_vision=False).
- [PR] docs: clarify that top-level modelContextLimit / ACP_MODEL_CONTEXT_LIMIT are a last-resort window fallback
- [PR] docs: How-it-works SVG diagrams (en/zh) replacing the ASCII box
- [PR] Review follow-ups: 429 throttle saturation/scope guardrails, media-refs carry-forward filter, oversize-refusal messages that say what to do, doc typo
- macOS desktop builds and full release notes coming soon.
- Added the run_workflow tool for everyone: the agent can run a Python workflow script that fans work out to many subagents in parallel or in stages and then combines their results.
- Added a pluginControls section to managed admin settings, so organizations that don't use Gemini Enterprise can control plugins too: pluginsDisabled stops all plugins from loading, including…
- Improved Auto mode's review of commands that would leave the terminal sandbox, so it decides faster and times out less often: the reviewer can now search the conversation, list directories, and…
- Cheaper than vector RAG: for both index and query.
- Faster: trees are created from layout stats.
- Unified SDK: vectorless RAG. Local or cloud, for index and chat.
- release(langgraph): 1.2.22
- release(checkpoint-postgres): 3.2.1
- release(prebuilt): 1.1.2
- Program agents: an agent can run your own code on every turn instead of a model, with a time limit and a payload input.
- Chats in My Kortix keep their own goals and check-ins until the work is done.
- A new workspace opens only after a model is connected, so the first chat always has a model to answer with.
- Feat(kubernetes): let the model target namespaces inside an operator allowlist (110607feb) — junji hashimoto
- feat(monaco): add Monaco CRM integration
- feat(computer-use): batch 2-8 actions per call and read app state in one round trip per element
- improvement(search): skip the Slack channel lookup on channel-scoped searches and read channel names in any script
- feat(db): label advisory-lock metrics with the locked event kind (359839f70b6ec2401207c9fb29116efc87b75701)
- perf(mobile): parse each event's thread reference once (bf422ac8e1a34893e15c91bfc86a74d990547feb)
- perf(mobile): load only the directory when search opens (c97c85d8faee7aee989f5a110e38cd38eff551b2)
- Fixed false MCP startup session.error events when quitting or switching sessions cancels in-progress MCP startup; genuine startup failures on live sessions are still reported.
- Fixed session.plugins.list incorrectly reporting user-installed plugins as disabled when an organization requires managed plugins.
- Fix taskAnalytics.classifyTrajectory retries rejecting an unchanged archive whose JSON object keys were stored in a different order. Results from earlier versions need one fresh analysis.
Nov 13, 2026
- A failed /compact now shows the model provider's error message without an internal host-rethrow prefix
- A prompt the CLI does not send yet, such as one entered while managed account policy is still loading, keeps its pasted text, so pressing Enter again sends the paste instead of the literal [Paste…
- The /plugin manager no longer shows a repository plugin that did not load as enabled, and plugin lists no longer say its directory was not found when its source points outside the marketplace
- Black frames at cuts. A clip split or trimmed between two frames no longer leaves a black frame at the cut, in the preview or in the exported file.
- Merge works on clips cut between frames, both halves of a split clip included.
- Black preview on Windows. When a PC's graphics device drops out, Concat lets it go, tries it once more, then steps down, and as a last resort draws the preview on the processor instead of going…
- Updated dependencies [23ff635]
- @ai-sdk/gateway@3.0.237
- This release adds server-side agent runs through Intelligence, richer Learning capture, and several fixes to chat rendering and tool-call handling.
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- feat(examples): add in-app guide example
- feat(examples): add new agent examples
- feat: add opt-in private network egress allowlists
- Studio: Drag and resize on an opaque preview start from one gesture-start read (6bdf9291647f3, #7657).
- Engine: Render text without ClearType colour fringes on Windows (273672ab13112, #7663).
- Core: A looping preview keeps playing when a scene reloads during a hold for data (b8f630c643b2b, #7659).
- exp-code 0.1.7: the open-source coding and cowork agent (agent/). The sandbox runner is attached as one self-contained file with its sha256.
- Approval prompts ask a clear question and name the action instead of showing raw tool details.
- Gmail watches keep their place after a temporary Composio error. App discovery shows clear limit and outage reasons.
- On managed networks, the local engine uses private CAs trusted by your operating system for Composio HTTPS. Certificate checks stay on.
- Python security relock: all integration uv.lock files (lock files only, no Python floor changes)
- BREAKING: raise requires-python for security fixes across integrations (stacked on #24039)
- Python security relock: all integration uv.lock files (lock files only, no Python floor changes)
- bash pip install claude-agent-sdk==0.2.182
- Choose Remove or Keep from the keyboard after closing a session. When Cmd+K (Alt+K on Windows and Linux) closes a session with unsaved work, the Keep button now shows its focus ring even if you…
- Team reviews run your whole review skill and cannot write files through git. A review skill that grants its own commands, such as a lane router, used to fail to load in a team review, so the…
- A board the browser stops drawing says what is happening and keeps trying. After graphics card resets, a browser can refuse graphics for the site: Chrome lifts that after 2 minutes, while Firefox…
- @claude-flow/memory 3.3.3 forwards the configured dimension to AgentDB and installs the configured embedder, rejecting vectors of the wrong length.
- The CLI memory bridge passes a local embedder when a non-default model is configured, and labels stored rows by the embedder that actually produced them.
- Stores pinned to the default 384-dimension model are unchanged.
- Gemini 4 Hydrogen
- The agent drops old screenshots in one batch once 20 tool results carry one, instead of on every new screenshot.
- BU_NAME callers get their own Chromium when Chrome can't be used.
- When a browser setup error may be fixed in a newer release, the CLI and MCP tell the agent to upgrade.
- Fixed a security issue in how zstd-compressed responses are decoded
- Added a spinner tip in VS Code and similar integrated terminals about re-running a notebook cell, shown once Claude has read or edited a Jupyter notebook
- Added a line under the error shown when your organization's policy blocks a request, saying how to continue: run /rewind and change the message, or run /clear
- safety: planted-instruction probes and the runner that checks what ran (78d20e7)
- voice: a published speech track carries a machine-readable AI-generated mark (c24b4c5)
- guardrails: image checks receive the guard's deadline signal (fdcab69)
- Azure: Fixed 400 on /v1/decisions for azure/ deployments
- Batches and files: Fixed file ownership checks missing other spellings of a Gemini file id, such as files/<id> or its URL
- MCP: Fixed sign-in loop ending in Authentication incomplete on servers that answer mcp_all with detailed scopes, broken since v1.107.0
- Set DEEPAGENTS_CODE=1 in the local shell environment.
- Replace unencodable characters in redirected output.
- 1425cf0 release(deepagents-code): 0.2.8
- c0211e9: The CLI now authenticates every request with your login's access token instead of a stored API key.
- e2b auth login opens (or The E2B_DOCS_BASE override is removed.
- Access tokens are refreshed automatically before they expire, including during long-running commands such as sandbox logs --follow, sandbox metrics --follow and an attached sandbox create session.
- chore(sdk): update composio-client to 2.1.32
- feat(multiagent): async agent builders; report content-filtered and refused children as tool errors (3d741b47d)
- feat(site): add /frogs page (9b38e2e9c)
- feat(context)!: offload conditions decide granularity (typescript) (50a391e05)
- 最低支持版本降到 Android 10(minSdk 30 → 29),Android 10 / 11 也能安装运行。
- 锁屏解锁后任务不再卡在「等你解锁手机」,改用 isDeviceLocked 判定(#590)。
- 试用技能不再一次失败就停用,Token 预算用完不算技能失败(#569)。
- Lokutor: Conv 2.0 STT and Versa 2.0 TTS across nine languages.
- xAI: xai/stt-2 through Inference, with word alignment for adaptive interruption.
- Google: Gemini 4 Carbon, Gemini 4 Hydrogen (also through Inference), and gemini-decision-1 through the Decisions API.
- App: Tools (Beta): the agent can look things up on the web when a task needs it (search tool use), with
- App: Phone calls and text messages (Beta, experimental): Call and Text templates. A call dials, speaks through
- App: intents improve themselves: when an intent's saved steps miss and the AI finishes the task, that run's
- #4875 4c72e1d Thanks @7Sageer! - Support optional search_url and feedback_url fields on custom registry (api.json) providers for WebSearch and /feedback.
- #4876 d532f4f Thanks @7Sageer! - Remove the text-only "No attachment" option from /feedback (text-only /feedback submissions are no longer available), show the current session's files before…
- #4885 ec35f9a Thanks @kimi-agent-bot! - Fix multi-second loading delays when reopening long sessions with many tool calls.
- Right-click menu for center tab strip tabs
- Pick the default shell for new terminals
- Add the native UI building blocks (native overhaul 2/4)
- Add Console mode selection for conversations, commands, or both, plus independent command.onOpenUri routing.
- Recover Console inputs after session creation or delivery failures, preserve undelivered questions, and show theme-aware red error notices without storing client errors in conversation history or…
- Open .abstract.md and .overview.md links at their parent directory's L0 and L1 preview. Fix file-only system directory listings by omitting retrieval tags.
- [Computer Use] Computer Use is now on by default and asks before accessing each app. Choose Always, Approve for session, or Reject, and manage saved approvals in Settings.
- [Skills] Added an "Export skills to external agents" switch in Settings → Skills.
- [Agent] Agents on DeepSeek models now show a short intent line above each tool call. Agent sessions keep tool steps expanded by default and fold repeated calls of the same kind.
- Verify recognizes a suite that was already red on base with no parseable test ids, and excuses it only when the head run fails in the identical kind: same exit code, no parseable ids, and the same…
- Every package that holds a file proving a test runner (go.mod, Cargo.toml, a test file) now gets that runner, not only the package of the first such file, so a multi-module repo whose nested…
- An edit inside an indented method now names its enclosing method, so the lean gate and the fail-then-pass proof see the changed symbol instead of "(no declared symbol)"…
- Nothing to do. The browser extension updates itself.
- Compact & Resume works in Projects that already hold chats started through the app.
- Goal chats recover. A Goal chat whose Continue failed before any tool call used to sit idle for good; the app now recovers it. A Continue after the app's own Stop stays owed until it is sent.
- pi-subagents 0.88.1. A permission prompt or other dialog that opens while the fleet inspector is up is no longer hidden behind it.
- Preserve JATS inline text. PubMed and Europe PMC full-text results keep words inside inline markup in their original order, including subscripts and superscripts, without joining adjacent paragraphs.
- Patched hono. Alpha Hub's MCP SDK now resolves hono 4.13.18 and @hono/node-server 2.1.8, which fix advisories published on November 12…
- Restart a self-hosted server from the server menu
- Keep the saved OpenCode key away from the free models
- Add an app icon picker to the mobile app
- Agent’s durable option now controls resumable streams only (not checkpoint-backed durable execution).
- Added resumable streams to every Agent, not just durable agents. An agent registered on Mastra now publishes each run on its stream topic and keeps a replay cache, so a client can reconnect after…
- Avoid blocking the event loop while loading instruction files from a trusted Git ref and while discovering tool-result instruction reminders. Keep missing and untrusted-ref behavior unchanged.
- Updated dependencies [03a62d4]
- @truefoundry/trueforge-core@0.185.1
- Fixed a CLI issue where the /usage activity graph wrapped in narrow terminals; it now fits correctly.
- The permission_prompt notification hook no longer runs for a permission request that Assisted Permissions approves without asking.
- Fix .NET shell authentication cancellation cleanup callbacks being lost when a request finishes before queued notifications run.
- 恢复 A 的原有规则校验,统一九种语言 README 的待更新与 Stars 文案。
- 补全 B4 主流程的画面审核参数及命令示例,保留独立文字和画面审核。
- 独立修复 B5 Windows 路径比较,保留素材、指纹与审核校验。
- #1710 0920ef5 Thanks @kdy1! - The agent tells the model when an action opened a tab in the background (The app opened a new tab (tab 2 of 2). The screen still shows tab 1.
- #1600 f5c9099 Thanks @okwasniewski! - The built-in agent can try an upload when the step names no file.
- #1541 11dea90 Thanks @samanyudas! - Fail e2e run when an app command exits unexpectedly after readiness.
- Connect ChatGPT and Claude right above the first-chat composer.
- A Slack workspace now runs in one Kortix workspace; agents move between workspaces by their own rules.
- Provider keys need no feature flag; "Pooled Provider Secrets" moves out of Feature flags.
- Undo stays until you're done with it: a message with Undo, after you dismiss a posting, archive an application, change a status or leave out part of a resume, closed after 6 seconds unless the…
- The chosen option stands out: in Board / Table, the dashboard's period, Faithful / Balanced / Expressive and New application's choices, the chosen option was only a slightly lighter pill.
- The current page in the sidebar: the current page looked like any item under the pointer, so hovering another one made two look current.
- The usage bars next to the model name show your Claude plan again, in Simple and Advanced, with the 5-hour and weekly figures at the top of the model menu. Codex, Grok and OpenCode Go keep theirs.
- A cloud computer that stops answering is no longer shown as ready. MausBot restarts it once when that is safe, tells you and the bot plainly, and the computer panel has Restart computer and…
- A cloud computer that fails to start can be reset from the same panel.
- 修复强制默认预设会覆盖手动选择的问题:移除 global.forceDefaultOpsPtc 开关,改为行详情页的一次性按钮写入默认预设;apply、重挂载与卸载都不再改写默认,旧 profile 保存的该键仍可加载但不再生效。 @T-Auto
- 修复 PTC 在关闭常驻工具与动态加载时仍注册 allowedTools 候选池的问题:没有 loader 的作用域不再声明这些 API,首轮提示词不再多出约 22.5k tokens,常驻与默认加载行为不变。 @T-Auto
- 修复 Native 搜索诊断丢失并统一模式契约(#867)。 @Haoming9527
- Memory and stability - the host no longer runs out of memory or stalls while a large batch of files is published, and closed tasks are released from memory.
- Resource use - fewer file scans, lighter usage tracking and fewer chat history writes, and a credential-check loop that could spike CPU is fixed.
- Large tasks - tasks with thousands of artifact versions load their versions again, and documents with many attachments can be compacted.
- feat(clickhouse): send query parameters in the request body by default
- feat(ai-gateway): add Google AI Studio as a gateway provider
- feat(ai-gateway): rename to Connections & Models and hide Models tab without connections
- 酒馆:新的角色扮演模式。 抽屉顶栏的「模式」键可以在工作、伙伴、酒馆之间切换,设置 › 界面 › 启动时进入可以选默认进哪个。兼容 SillyTavern 的角色卡、预设、世界书、正则和聊天记录:导入卡(文件、分享、卡站链接)或自己写一张;对话有气泡、小说、视觉小说三种排版,支持备选、接着写、帮我写、分支和书签;资源库里管理预设、世界书、正则、人设和快速回复。消息里的 HTML…
- 创意工坊。 宠物设置和伙伴小屋的菜单里多了「创意工坊」:浏览、安装大家做的像素宠物和伙伴,不用登录;也可以投稿自己做的,先审后发,有问题的作品可以举报。投稿、查看自己的投稿和举报要用论坛账号,第一次会先说明会把什么发给工坊服务,同意后才发。
- 悬浮伙伴。 宠物的「App 外显示」可以选「小屋里的她」,让伙伴小屋的角色待在悬浮窗里。悬浮窗里点一下是互动,连点两下才打开对话(像素宠物也改成这样);伙伴小屋会跟着手机倾斜和摇晃动。悬浮伙伴比像素宠物更耗电。
- feat(manifest): record excluded files in run coverage
- fix(action): show warnings in zero-finding PR summaries
- fix(deps): update compatible vulnerable dependencies
- core: Stop redacting executions that use global credentials (c067170)
- 4dc251a47d597a4d974e286935a6d64c219e7bff EDEN-20: add confirmed deletion of inbox entries (@happice0746)
- a0371ab2cd615133871d9d66fd14fa29010bf07d MUL-3806: include directly requested agent issues in Created (@Bohan-J)
- e6c3ef3d8025064c37e4346214e53c636dce2171 MUL-7031 feat(issues): show requester beneath agent creator (@Bohan-J and @xiaoke12345678)
- Purge API: clear the whole memory root, one app or one project while the server keeps serving. No more stopping the server to delete files.
- POST /api/v2/memory/purge: start a purge.
- POST /api/v2/memory/purge/confirm: confirm it before anything is deleted, so nothing goes by mistake.
- Background Subagents by Default: Subagents now return immediately, continue in the background, and report back whenever they become idle.
- Quieter Transcripts by Default: The TUI now summarizes each run of tool calls and thoughts into one collapsible activity row, so the answer is easier to follow.
- Jump to an Earlier Prompt: In the desktop app, open Jump to message from the command palette or press Cmd+Alt+G (Ctrl+Alt+G on Windows and Linux) to search prompts across the current session and…
- agent-framework-core: Expose loop_return_final_only on create_harness_agent() so non-streaming callers can request only the final loop response without changing the existing default or streaming…
- agent-framework-core: Add the optional experimental SecureAgentConfig.message_labeler callback for application-owned classification of context-provider messages
- agent-framework-github-copilot: Support Pydantic models and JSON Schema mappings through response_format, including parsed response values in streaming and non-streaming runs
- fix(provider): stop rewriting adapter npm for mimo-named model ids
- fix(cli): unpack with the Windows system tar, not the first tar on PATH
- test(provider): cover Anthropic routing for MiMo models
- feat(demo): auto-confirm demo bookings that came through the verified form and decline the rest
- improvement(desktop-browser): let a page finish rendering before observing an action
- improvement(search): shorter Jira any-term budget and a shared batched read budget
- ⚡ Automations can now be triggered by Schedule, Jira, Slack, GitHub, Linear or PagerDuty, chosen right in the “New automation” form.
- ⌨️ The desktop app has a native menu and a full set of shortcuts: new terminal and browser tabs, pin, archive, fork, sidebar toggles, switching workspace tabs, reopening closed tabs, back and…
- 🔍 The pull request Changes tab can now filter changed files, find text in the diff with Cmd/Ctrl+F, mark files as viewed and track progress, and browse each commit's diff…
- app: add pnpm dev:ui-only and pnpm dev:down (4ab080c)
- cli: add dataset-versions list command (1e47750), closes
- cli: add prompt-versions list command (974be43)
No release notes published.
- Added Gemini 4 Hydrogen (Low, Medium, High, and XHigh) and Gemini 4 Carbon (Low, Medium, and High) to /model and --model when you connect with a GEMINI_API_KEY, added Gemini 4 Hydrogen…
- Added click-to-expand in the full screen (altscreen) rendering mode: clicking a tool group, a thought, a prompt's attached images, or multi-line command output now expands or collapses just that…
- Added context window usage updates to agy acp sessions, so editors such as JetBrains AI Chat can show how much of the model's context window the conversation is using.
- Plugins, first version. SolidCraft can run sandboxed WebAssembly plugins. A plugin works only through SolidCraft's own commands, the same ones scripts and AI agents use, and only with the…
- Install a plugin from its .zip with Install plugin… in Preferences › Plugins or the Plugins menu. It shows what the plugin asks for before installing.
- A Plugins menu runs installed plugins; each one gets a dialog built from its parameters.
- Browse model providers under Sessions → Model providers and choose the default for new chats.
- Enable features.tui_workspace to view a side conversation beside the main chat. Drag the divider to resize the panes or press leader z to zoom the focused pane.
- Attach recordings when using an audio-capable model. Codex skips confirmed rejected recordings on later turns and uses clip duration to keep more short recordings in context.
- commit: 7227e8311a2d782f399323b21483c072fb1353b8
- source tag: beta2026-11-13_10-54-29
- regions: cn global
- Dynamic credentials: Credential Vault Exchange resolves credentials per request across the server, egress, and all five SDKs, without vault persistence of resolved secrets.
- Multi-version Fast Sandbox VMMs: managed Firecracker 1.16.1/1.17.0 bundles let new template builds select the highest installed version while existing artifacts restore with their exact recorded…
- Safer data and capacity management: staged uploads reduce overwrite data loss where atomic replacement is available, controller restarts preserve live-sandbox Pods, and Docker enforces requested…
- 📱 Scroll phone chats again: On chat pages served by the desktop, swipe through messages and sideways through wide code blocks.
- 💬 Use GitHub Copilot in Chat: Choose Launch chat to stream replies, inspect tool output and review file diffs in a conversation you can reopen after restarting Orca.
- ⌨️ Start Pi with your chosen model and thinking level: The terminal launcher applies your model choice and offers reasoning effort before you send the first prompt.
- fix(ci): read prettier's whole ignore set at the base tip
- fix(cli,core): let an Agent Host retry an enrollment whose answer was lost
- fix(channels): shape-check authorExclude on the inbound task record
- docs: bring the process entrypoint map in line with current hosts (1ef6ed233) — vaibhav upreti
- fix(guardrails): validate positive limit in audit command and return empty on non-positive limit (f5e3261d3) — Sarthak Chaudhari
Nov 12, 2026
- prepareCompaction() runs as the compaction.prepare loop phase, so a ui.loop-blocked stall from a cold token walk over a long branch names compaction instead of "unknown"
- Added the google-decisions judgment API for Google's Gemini Decisions API (gemini-decision-1), with text and image input
- Added apiKey to successful OAuthAccessResolution results from oauth.accessById() and oauth.accessAll(): the key normal requests send for that account, which differs from accessToken for providers…
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Fix managed-account changes stalling while live extensions restart and rejoin the session.
- Copilot SDKs can now answer sudo, SSH, and Git password prompts from commands the agent runs, through a new shell authentication handler. Answers go only to the program that asked.
- Node.js, Python, .NET, Java, Rust: A request handler no longer keeps memory for each response the runtime cancels after the handler has finished it, such as a response rejected for exceeding the…
- bash pip install claude-agent-sdk==0.2.181
- Added snapshot hints for controls whose accessible name is empty or shared with another ref in the same snapshot.
- Added a Modern Web Guidance skill that retrieves Chrome's guidance for a task and checks the built page in agent-browser
- Fixed commands on closed tabs waiting for the timeout. A command whose tab closes before Chrome answers now fails with Target closed: CDP session <id> detached
- Experimental: answer git, ssh, and sudo password prompts from agent commands in a masked dialog (enable with /experimental on)
- Avoid unnecessary error popups during Microsoft Entra sign-in when an organization account is available
- Screen readers identify the current session and announce when sessions need attention or are idle in the Sessions list
- feat(site): add /frogs page (9b38e2e9c)
- feat(context)!: offload conditions decide granularity (typescript) (50a391e05)
- feat(mcp): add find_integrations tool backed by the catalog feed (085d4de61)
- Added a fully-featured Slack connector and @cognee Slack app so Slack messages can be asked and remembered directly from Slack.
- Introduced an embedding-worker engine over RabbitMQ so background embedding generation is more robust and scalable.
- Improved Google-family connectors (Drive, Gmail, Calendar) including handling legacy file formats and retry logic so fewer files are skipped.
- Amazon Q for VS Code 2.10.0
- Core: Runtime element handles for the editing channel (e1ddb2e16, #7603).
- CLI: Check hears a WAV whose header misstates its byte rate, as render plays it (97df93fb0, #7618).
- Engine: A frame's pixels no longer depend on the frames drawn before it (b4fc4736d, #7181).
- Added Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools; see to turn it on, or to turn it off for an organization
- Added engines["claude-code"] to plugin.json: plugins outside their stated Claude Code version range are not installed, updated or loaded, and install and update pick the newest version that fits
- Added type checking of a plugin's TypeScript hooks module to claude plugin validate; each finding is a warning, and --strict fails on it
- neural wsp — optional ruv-FANN weight-space tooling (ADR-528, #4980). ruflo neural wsp status | inspect | check | gate | export-centroid. The three ruv-fann-wsp* packages are optionalDependencies.
- Opt-in model-judge routing in ruflo-mods 0.17.0 (#4969, ADR-530). routeJudge, off by default: one haiku label per prompt, any failure keeps the keyword route.
- Windows: the remaining 2>/dev/null npm redirects removed, and analyze deps --security now reports when npm audit exits non-zero
- Fix: csharp a member call on a receiver typed by a lowercase, underscore-prefixed or keyword-like type (_Pair, pair, a field or primary-constructor parameter so typed) now binds, and a static call…
- Fix: java a call through a class-qualified static field (Service.INSTANCE.get(), and a chain Service.INSTANCE.other().m()) resolves via that field's declared type.
- Fix: sql a routine body preserves comma joins (FROM a, b) so every table is read, and keeps a @dblink suffix as a distinct remote node instead of pointing the edge at a local same-named table
- tracing: turn off FastAPI's native tracing on Agent Engine (87e5399)
- Context meter after a reload or model switch. For some sessions on custom or self-hosted endpoints, the
- Credentials can pair with the wrong URL on Agent builds after hermes-agent #151775 when a providers: entry named like a built-in provider has its own URL and model.base_url differs (#8973, Form 3).
- release: #8903 TestConflictError no longer fails in focused or partial test runs
- This patch release fixes A2UI rendering for catalogs that declare protocol version 1.0 when streaming Direct JSON.
- Gemini 4 Hydrogen works through the Gemini provider, including streamed replies,
- The chat widget shows Markdown tables in an agent's reply as tables. They
- Event streams (sse=true) now carry an id on every frame and resume from
- Responses API: Set drop_params: true or stop sending Responses-only fields such as max_tool_calls to providers bridged through chat completions <sup>Also in 1.109.0</sup>
- Azure: Added gpt-image edits on Azure AI Foundry deployments, including deployments under a custom name
- AWS Bedrock: Added image search on Bedrock knowledge bases with query_image, including image hits in the results
- 11323f9: Prevent excessive memory and CPU usage when validating malformed UI messages.
- 2ebc78b: Forward the provider's configured fetch transport when downloading file inputs.
- cb4fa52: add experimental extract api and extraction model spec
- Cache optional prompt-caching provider lookups in the SDK.
- 0ec4244 release(deepagents): 0.7.37
- 4a51d63 fix(sdk): cache optional prompt caching provider lookups
- feat(ace): expose typed page tasks and browser runtime hooks
- fix(dump): nest and strip child frame documents in html dumps
- fix(dump): include and strip shadow roots in html dumps
- Codex keeps recording after claude-mem updates
- Optional local API token for the worker API and MCP HTTP transport
- The quota fallback warning names both providers
- Allow approved private URL hosts.
- Fix projects to depend on PyMuPDF 2.x.
- Drop a real storage URL from the URL read tests.
- Updates: From the next update on, the update notice links to its release notes, so you can see what changed before you restart.
- Quiet hours: From and Until follow your 12-hour or 24-hour clock setting.
- Memory: Notes with tweet, LinkedIn or order links save normally. Long IDs no longer look like card numbers.
- release(langgraph): 1.2.21
- chore(deps): bump the minor-and-patch group across 1 directory with 9 updates
- fix(langgraph): match whole namespace segments in get_subgraphs
No release notes published.
- Prevent duplicate Skill Recorder instances. Electron's native single-instance lock is acquired before recorder initialization.
- Restore the existing window. Repeated launches restore a minimized recorder; requests received during startup are handled after initialization.
- Make Windows upgrades explicit. New installs and same-commit reinstalls both finish with manual launch instructions rather than automatically opening another copy.
- OS Manager defaults and validation: OSManager.tool_call_limit now defaults to 24 (was 8), and num_history_runs defaults to 5 (was 7).
- A2A v1.0: AgentOS and A2AClient now run on the official a2a-sdk 1.x, so upgrade the Agno server and client together.
- The /v1/message:* and /v1/tasks:* routes are removed; use each entity's A2A endpoint instead.
- Sessions can wake themselves later. Claude Code, Codex and Kimi Code sessions can run glimmervoid wake in 15m <text>, glimmervoid wake at 14:30 <text> or glimmervoid wake cron "0 9 * * 1-5"…
- Resize the width of a board card you open full screen. A picture, video or HTML page opened from a whiteboard now has a handle on its right edge: drag it to narrow or widen the card, and an HTML…
- Wide HTML pages fit the full-screen board viewer. A page whose content, such as a wide table, is wider than the viewer now opens scaled down so the whole thing shows, with no sideways scrolling…
- feat(core): drop the tool round cap (c415b56), closes
- In the Android phone app, tapping another Bot or a group chat in the sidebar or tapping an incoming notice banner no longer closes the app, and switching chats keeps each conversation's cached…
- In the Android phone app, opening Memory for a Bot or a group chat that has a saved summary no longer closes the app.
- The Android APK is signed with v1 (JAR), v2 and v3 signature schemes together so OEM package installers recognize its developer certificate on first install.
- cli-mcp exits cleanly when the client closed the pipe before the reply (was exit 1 with ExceptionGroup).
- The CLI runs its own Chromium when no browser has remote debugging on or there is no display, like --cli-mcp. It persists across calls; browser-use --reload closes it.
- A BU_CDP_URL nothing listens on fails in 3 s with one message naming the value and both ways out (was a 30 s wait).
No release notes published.
- 适配 DSH 0.2.2-rc.2 — 主适配线前移到 0.2.2-rc.2,peer 范围同时保留 0.2.2-rc.1、0.2.1-rc.5 等旧验证线,停在上一验证线的用户也能直接装这版。rc.2 上 Agent Teams 改由 dsh-base 内置、按会话开关,成员改用具名子代理;账号通知改用新的通知…
- 通知覆盖所有等待状态 — 以前只有任务完成会发系统通知;现在等待工具授权、等待回答的提问与计划评审也会提醒,同一个「通知」开关管四类通知(/settings → 对话:taskNotifications 默认 unfocused 仅失焦,可选 always /…
- 点击通知回到发起它的窗格 — Windows Terminal 上开启 jumpBack(/settings → 对话,默认关闭)后,完成、审批、提问与计划评审的 toast 都带「跳回」目标:发送前先做一次只读探测,恰好命中一个窗格或标签标题时才承诺跳回;点击时再核验五项(同一窗格、获得键盘焦点、标题仍含标记、窗口在前台、未最小化),任一步不成立就不动作。无法核验的布局(非…
- bsk open, bsk sync, and bsk close manage a browser session for the current caller. Use --scope <name> to isolate parallel tasks. Sync returns short context and observation IDs for subsequent actions.
- bsk plan runs a sequence of decided actions and stops on the first incomplete step.
- bsk read samples bounded DOM fields such as text, attributes, and form state on changing pages. MCP and DSH expose the same read operation.
- 2b7816fed084f029f4061e441a6f9f3980e71c21: feat(skill): toggle skills (@BrunoKrugel)
- 7f9e4824c465e1ffa1a1d652dd6e9a897b54035c: feat(ui): Add elapsed time on tool call (@BrunoKrugel)
- f38c0dc030912cc3eeaa82cfb014cf21868405b0: feat: add CRUSH_DATA_DIR env var to override the project data directory (@andreynering)
- Changed the default Agent Mode from Default to Auto: sessions with no saved mode, including headless -p runs, now start in Auto wherever Auto mode is available and in Default elsewhere.
- Fixed multi-second UI freezes, which also stalled streaming and window resizing, when a command printed a very long line without spaces; line wrapping now takes time proportional to the line's length.
- Fixed agent replies stalling for 20 seconds or more while a group of tool calls with long output was still animating its token counts; streamed steps now keep appearing while the group re-renders.
- [Tasks] Remote clients can discover scheduled-task result conversations and query task schedules, run history, and bounded summaries.
- [Models] Rapid-MLX is available as a built-in, keyless local provider.
- [Web Search] Added SerpKite and StepFun keyword search providers, and Exa can now fetch the content of a URL through its Contents API.
- integrations: add Approve and Skip buttons to Google Chat cards (9f0c9dc)
- ui: give email cards one set of buttons and answer questions on them (33073d0)
- delegation: reply in the thread to whoever they ask, on the waiting card when there is one (5df30ac)
- Open the sign-in view on a server after a saved login fails
- Show a member only the Library files of the agents that they can see
- Find the Gemini verification link in each text format
- Updated third-party dependencies with known vulnerabilities, including the HTTP client used by model, mail and scraping connections, the mail sender, and XML and Markdown parsers.
- The development launcher's database cleanup accepts only database names it generated and quotes them, so forged database metadata can no longer inject SQL or point the cleanup at another folder.
- On Apple silicon Macs running macOS 26 or newer, the desktop app's This computer runs without Docker.
- [Workflow: CC] Marking a workflow CC task as read or unread now preserves its recipient and linked workflow records.
- [Custom variables] Creating, editing and deleting custom variables now requires the “Allows to configure interface” permission.
- [Authentication]
- feat: improve Codex subscription authentication guidance
- fix(chat): keep messages at tied history page boundaries
- fix(chat): prevent delayed echoes from deleting newer drafts
- fix(workspaces): create worktrees in the background so a large checkout never freezes the app
- chore(github): remove SBOne-Kenobi from CODEOWNERS
- fix(web): highlight every catalog language in chat
- New JuiceFS COW storage backend (preview): built on JuiceFS, it enables sharing templates across nodes, resuming a paused sandbox on a different node, and booting from a snapshot on a different node.
- New FS-only snapshot capability: snapshot only the sandbox filesystem to cut storage costs.
- New sandbox restart policies: automatically restart a sandbox when the service inside it crashes.
- #2142 5efdeb3 - Add the Muse provider adapter. It drives Muse's OpenAI-compatible Chat Completions endpoint and ships the Muse model family, including the flagship muse-ember-1.0 (Ember) model.
- fix(deps): patch vulnerable chatbot test toolchain
- ci: mirror the platform image to TCR after publishing GHCR
- fix(runtime): keep index.lock while an alternate-path holder is live
- Commands report their own lifecycle — every command emits a start and end event, including run, testmd run and explore, so a run's shape is visible even if it dies early.
- Sessions say what they are and how they ended — session start records the surface and target; session end records the outcome, including sessions that exit without finishing.
- Failures carry an error code, a domain and the actual error — no more guessing from a bare stack trace, and a testmd run that throws now logs why before closing its session.
- 支持上下文使用次数与最近使用时间标签更新
- 支持批量设置上下文标签
- 支持 Python 同步与异步客户端调用 batch_set_tags
- memory: a souled agent's memory and wiki pages are kept per user, in a new conversation scope.
- guardrails: a guard can ask to run first, and a closed failure names its cause (142814d)
- guardrails: a typed moderation context (917f1dd)
- Python 3.10 or newer is required. Python 3.9 is no longer supported; native wheels use the Python 3.10+ ABI.
- The next mine re-reads each source it covers once. Normalization version 3 picks up the parsing and byte-preservation fixes. Drawers whose text is unchanged reuse their stored vectors.
- Hallways migrate to SQLite on the first hallway write. hallways.json is imported transactionally into hallways.sqlite3 and retained under an .imported-* name.
- commit: 94475474ec60d13f94790d0548e74b7d75ce9de6
- source tag: beta2026-11-12_17-55-23
- regions: cn global
- #4752 5aa81c5 Thanks @sailist! - Align fs:suggest with local fd completion, using bounded scans and at most 20 results with optional streamed previews, while including hidden and gitignored paths…
- #4618 e6ae8ed Thanks @Grapedge! - Fold consecutive tool calls and thinking into summary lines in fullscreen mode, and no longer show the Activated skill card for skills the model loads.
- #4751 b64a016 Thanks @Grapedge! - Show how long the current turn has been running in the activity bar while the model writes, and how long the last turn took and when it finished once it completes.
- fix(xlsx): place a new dxfs before tableStyles and colors
- fix(pptx): give a materialised placeholder a unique id
- fix(xlsx): wrap English text without breaking words
- feat(ai-gateway): add OpenRouter as a pre-release gateway provider
- feat: add SUBSTITUTE_OK no-data mode for monitors
- feat(ai-gateway): add xAI as a pre-release gateway provider
- The executable-spec call is off by default (LOKI_E10_EXEC_SPEC=1 opts back in). On the 20-task Haiku tuning bench it never became a binding check, and since v11.137.4 one call could carry up to…
- The separate coverage turn is folded into the implement brief: the implementer is asked up front to change and test every site the task names (LOKI_E10_COVERAGE_RESUME=1 restores the separate turn).
- The Opus consult stays off by default as in v11.138.1, now grouped with the other lean defaults; LOKI_E10_LEAN_CALLS=0 restores every v11.138.0 default at once.
- <!-- 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)"…
- CLI archives built from the Studio release studio-v2.61.0.
- 🔴 The phone will not tap WeChat, and that is a protection rather than a limit. One automated tap puts
- You can quit without a balance. The control bar's ✕ was minimise, and typing 退出 went to the model —
- A mic switch that keeps your enrolment. Wake listening held the microphone 24/7 for people who only
- Connect Gmail and other apps from the chat: a Connect card appears when the agent needs a sign-in, and the connection finishes by itself when you come back.
- Provider keys are checked when you save them. Models → Providers shows whether the provider accepted each key, with "Check again".
- OpenAI API keys can be added in Models → Providers, next to ChatGPT accounts.
- fix(cal): read action results from the Cal.com data envelope
- fix(metabase): pass native MCP tools through
- perf(mcp)!: bound discovery results and defer schema loading
- Slides: legacy PowerPoint 97-2003 .ppt files now open — from the Open dialog, drag and drop, file association or the CLI — as an editable .pptx copy (@tokg-venv).
- Docs: Save As offers PDF, HTML and Markdown next to Word; a story view hides the formatting while you write (@tokg-venv); update the table of contents from its context menu (@aniruddhaadak80).
- Sheets: slicers and timelines persist across save (@tokg-venv); break and change-source for external links with a cross-workbook range picker (@tokg-venv).
- feat(tools): support reading both comparison sides
- feat: add GitHub Copilot CLI integration (plugin manifest + guide)
- feat(llm): prefer MiniMax M3.1 Flash in model suggestions
- Mika-v0.1.40-arm64.apk:近几年的安卓手机基本都选这个,更小。
- Mika-v0.1.40.apk:通用包,所有手机都能装。不确定就选它。
- 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
- Copy and paste work again on the desktop. Since 0.40.0, copying did nothing in the desktop app: Copy Error Message, Copy diagnostic information, Ctrl+C in a text field, and paste.
- A cut through a body's edge builds again when its sketch sits a few micrometres off that edge. It used to fail with "SMLib's check refuses the result: SmFace::Test #36 Face Near missing Seam".
- Lofts through parallel sections no longer dip behind their first section, so a carved top joins onto the slab it starts from again, as it did in 0.30.1.
- feat(aws-marketplace): gate Enterprise on the AWS Marketplace container license and add the EKS install profile
- feat(credentials): audit fixes and non-OAuth auth for CrowdStrike, Datadog, Gong, PlanetScale, SailPoint, ServiceNow, Zendesk, ZoomInfo, Shopify, Atlassian, Linear and Twilio
- feat(api): failing block on polled runs, deployed input schema, paused-runs list, and applied limits
- GitHub Action for Codex Security 0.10.0.
- Installs @openai/codex-security@0.10.0.
- yaml uses: openai/codex-security@3bde66eda1a5c6b0c4cd245f51c16424c9374afb # action-v0.10.0
- Fix typing delays in answer and thinking fields during lessons.
- Fix Flashcards generation failing on OpenCode Go with MissingSessionID.
- Fix compaction overwriting newly saved skips, question dismissals, direction changes and pause state.
- Dashboard sign-in through your identity provider. It's opt-in and off by default.
- [gateway.dashboard_sign_in] offers one or more [oidc.<alias>] providers on the dashboard's sign-in screen.
- The access token stays in the gateway's memory, and the browser holds only an HttpOnly session cookie.
- Note text that looks like an application frame in user and tool content instead of escaping it
- Add OutputTokenLimitExceeded, and fall back to the next model in ModelFallback when a response is cut off at max_tokens
- Ask for approvals and answers inside the run with RequestInput, put AskUser on the deferred pipeline, and mark deferred calls with kind and answered_by
- 🌐 Keep your browser feedback across restarts: Unsent text notes on pages in Orca’s browser now return after a reload or restart.
- 🎙️ Give yourself time to think while dictating: In Settings → Voice → Input → Stop After Silence, choose 30 seconds, one minute, two minutes or Off.
- ⌨️ Paste while a terminal connects: A command pasted into a starting or reconnecting terminal now waits for its shell instead of being dropped.
- exp-code 0.1.2: the open-source coding and cowork agent (agent/). The sandbox runner is attached as one self-contained file with its sha256.
- refactor: remove the bootstrap package (38c575624) — Vaibhav Upreti
- feat(onboarding): explain each startup menu option (e99b9e8b1) — Vaibhav Upreti
- chore: enforce package layers and gateway borders from one import-linter config (f9e505129) — vaibhav upreti
- feat: authenticate gh on agent machines with Git credentials
- fix: return path_not_found for missing memory paths
- Added a normalization field to CartesiaTTSService.Settings and CartesiaHttpTTSService.Settings, passing Cartesia's text normalization mode ("auto", "off", or a locale code such as "en-GB") through…
- Added SarvamHttpSTTService, which sends each VAD speech segment to Sarvam's REST speech-to-text API (saaras:v4 by default, or saaras:v3) and yields the transcription from run_stt().
- Added max_speakers and filter_profanity to AssemblyAISTTService.Settings, passing AssemblyAI's speaker-label cap and profanity filter through to the streaming session.
- Three groups, always in the same order: the place you are in (This computer or a Cloud), Your account, and This app. Every page says what it applies to.
- Simple mode shows fewer pages. Turn on Advanced mode at the bottom of Settings to see the rest.
- One AI page for engines, provider keys and models.
Nov 11, 2026
- The optional agenticow peer is now pinned exactly to 0.3.1 (was the open range ^0.2.0, which excluded the current agenticow 0.3.x).
- Upgrade note: a project that has agenticow 0.2.x installed alongside this package will now see a peer conflict. Install agenticow@0.3.1.
- New __tests__/agenticow-live.test.ts imports the real agenticow@0.3.1 (exact devDependency) and drives AgenticowMemoryProvider + AgenticowQueryProvider through branch, ingest, native cross-branch…
- This release adds support for rendering A2UI 1.0 operations alongside the existing v0.9 protocol, so agents emitting native 1.0 surfaces (such as those from ADK) now render correctly without…
- Move agent-run turn helpers into focused modules
- Turn on ALB access logs in the AWS scaffold
- Keep HTTP keep-alive sockets open longer than the ALB idle timeout
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Verified findings (S1): findings carry a POE level and a verification status. verified requires a bound proof or verification response, a declared POE level of 3 or more, and every evidence quote…
- Target gate (S3b phase 1 and step 2.1): opt-in per task or conversation, API only, OFF by default. Approvals pin host:port and the resolved IPs.
- Agentic hardening: code-owned untrusted-content rules for the planner and the compactor, a planner goal anchor, goal_met and prove_goal guards, truthful hint origins, and the [built-in] deny rules…
- Codex now sends tool updates incrementally by default on Google Cloud Vertex AI. To opt out, set incremental_tools_legacy_gcp = false under [features] in config.toml.
- Security: Preview bundles never include .env files or git data, even through a link. A render that names one of these files now stops with an error.
- Security: Helper programs are found by walking your trusted PATH folders directly. On Windows, the working folder is never searched (9a048ac01, #7562).
- Share: Change detection keeps working when .hyperframesignore uses gitignore patterns such as /exports/ or **/*.mov. Before, those patterns stopped it with an error (c07468d6a, #7568).
- Remove chromadb from the embedding function layer.
- Add exact historical state lookup by row id.
- Wire jsonl into rag data type discovery.
- integrations: treat a Google Chat sender on People as signed in (c2a4e41)
- chat: log a timed-out suggestion call as a warning (e87b0ff)
- chat: put held steps on one Approve card and keep draft files (80d3889)
- opencode-pty 0.5.2 (protocol 8) from packages/process-daemon, published to npm as @opencode/pty@0.5.2.
- memory: send recalled text as a data block in a user message, not as a system message
- memory: run the memory hooks once per turn on agent(), and log a failing hook as an event
- guardrails: a guard lists the codes it can answer (e406065)
No release notes published.
- Fix: cpp a folded parameterless member declaration (bar()) records its implementation in definition_file as a path anchored to the same root as source_file, instead of a ../-walk relative to the…
- Fix: llm a slice whose timeout or context overflow is recognised and given up on (it cannot be split further, or the retry depth is exhausted) now marks its file partial, so a document whose tail…
- Fix: php a route path scan stops at the closure's own argument, so $app->get('/users', function () { ...
- exp-code 0.1.0: the open-source coding and cowork agent (agent/). The sandbox runner is attached as one self-contained file with its sha256.
- Added an edit_predictions.button setting to toggle the edit prediction button in the status bar. (#66450; thanks z0al)
- Improved compatibility with the latest Agent Client Protocol, including handling of missing sessions, plan updates, diff formats, and cancellation of permission requests.
- Improved the Agent's read_file tool output by truncating lines longer than 2,000 characters to prevent context window overflow. (#65395; thanks Antphar)
- Tool calls get ctx.toolCall.earlierEffectsSettled, which settles once the earlier calls of their message that run beside them and may write files…
- Added SOFYA_API_KEY as the env key for the sofya provider
- Devin requests no longer send temperature: 0: the catalog's temperature-floor raises it to 0.0001, so session titles and Mnemopi memory extraction work on Devin instead of failing with Cascade's…
- ChatGPT and Claude subscription usage lives in the OpenAI and Anthropic provider sheets: every connected account, every usage window the provider reports, and when the next login is back.
- The mobile app lists, opens and shares a project's Artifacts.
- Mobile app 1.6.1 on iOS: native system sheets, a ChatGPT-style project drawer, and a full-screen GenUI map.
- 91a67c2: Fix excessive CPU usage when streaming long arrays with Output.array or streamObject.
- 84b1cb0: Cancel pending send-now operations when a later stop or dispose occurs, preventing an abandoned chat's queued messages and history from being sent through a replacement transport.
- Updated dependencies [eaeb37e]
- #2136 701ec9c - Checkpoint a sandbox workspace at any time, and restore it after a crash.
- @tanstack/ai-sandbox
- New captureSandboxCheckpoint, restoreSandboxCheckpoint and sweepSandboxBlobs. They work on any SandboxHandle, also with sandboxWorkspaceBackend.
- Context: an explicitly empty project selection shows nothing instead of observations from every project
- Context: read-cost and savings estimates count stored facts once, instead of counting extra escape characters and an invented [] for rows without facts
- Observer: a streamed answer that arrives as a refusal counts as generated output, so an interrupted stream is not sent again and charged twice
- Added the swarm's queen bee to the /swarm board header and a 🐝 beside each live member in the Roster.
- Added a continuation_active field to the JSON input of Stop hooks, which is true while the agent is already working on a continuation an earlier Stop hook requested, so a hook such as "run the…
- Added a supportsThoughtCirculation option to admin-managed model entries (models.list[] in admin settings), so organizations can let non-Gemini models behind an enterprise model gateway reuse…
- Updated the IP-address parser to fix a regular-expression denial of service affecting address parsing in rate limiting and proxy connections.
- Failed MCP responses now have a total size limit even when they use an event stream, preventing oversized error responses from being buffered without a bound.
- In the desktop app's This computer mode, bots can see and use Mac and Windows apps when Let bots use this computer is enabled.
- 重新打开已有环境。 基础模板缺失或安装未完成时,已有虚拟机仍可从自己的磁盘启动。创建或重置环境需要可用模板,重置会在停止或删除原环境之前完成这项检查。
- 下载大工具包。 支持分段的服务器最多可用 4 个连接并行下载。中断后可以复用已保存的进度;服务器不支持分段时自动回退为单连接。下载完成后仍会校验整个文件。实际提速幅度取决于线路和服务器限速。
- 在终端使用 GitHub。 新环境默认安装 GitHub CLI,已有环境更新工具后也可使用。在 Cage 终端运行 gh auth login,用自己的 GitHub 账号授权后即可按仓库权限处理提交、PR 和 Issue。终端使用说明
- feat(automations): filter the dashboard by owner with a Service accounts option
- feat(chat): move the message hover timestamp closer to its message
- feat(conversation): keep Stop hook feedback out of the chat
- Stop the Computer Use driver after the agent service at quit
- Fail passkey requests at once in the macOS built-in browser
- Stop the Windows agent dock from slowing the whole computer
- Hosted servers read GitHub through a GitHub App, with no personal token. A hosted doorman with the Glimmervoid GitHub App configured mints read-only tokens limited to your installed repositories…
- Delete a board from the Focus rail. Each board row now has the same remove button a session row has. Pressing it dims the row and gives you Undo with the same countdown as closing a session.
- A hosted server asks to be put to sleep when nothing is working. Once no session has been starting, working, running background work or waiting on a scheduled wakeup for 20 minutes of awake time…
- fix(sop): provide a default sandbox and validate sandbox selection
- chore(release): prepare v0.3.6
- ci: keep browser jobs running when runners cannot reach the Ubuntu archives
- langchain 1.5.0 drops support for Python 3.10, which has reached end-of-life. The minimum supported version is now Python 3.11, and the minimum langchain-core version is now 1.8.0.
- Changes since langchain==1.4.10
- release(langchain): 1.5.0 fix(langchain): mark ModelRouterMiddleware as beta feat(langchain): Remove EOL'd Python 3.10 from langchain_v1 feat(langchain): add model router and auto mode middleware…
- DS Bot phone app for Android (beta). Pair a phone with the QR code in Settings → Phone, then chat with your Bots and group chats, send photos, answer questions, approvals and cards, and manage…
- Six more interface languages: Traditional Chinese, Japanese, Korean, Spanish, French and German, in the desktop app, the plugin and the start window.
- Group chat messages, the calls a Bot makes to other Bots in a group, and messages between Bots are written to a ledger before they go out, and a group's round now runs inside DS Bot.
- Nearest-centroid agent router (ADR-527, #4953). Set CLAUDE_FLOW_ROUTER_CENTROID=1 to route tasks to agents by nearest centroid over labelled examples.
- Neural router assets now ship in the npm package (#4958, fixes #4954). With CLAUDE_FLOW_ROUTER_NEURAL=1, an npm install now loads the bundled KRR model, the low/med/high bucket specialists and…
- darwin pin moved to ~0.14.0 (#4956, fixes #4952). This fixes the clean-install and pin-drift CI failures. Without a surrogate config, darwin 0.14.0 behaves as 0.13.1.
- pi-subagents 0.88.0. A child waiting for your answer no longer runs out its timeoutMs while it waits, and a reply cut off at the model's output limit now fails instead of passing as a success.
- Paper search accepts sort: "citation". feynman_science_database_search recommends Semantic Scholar's default citation-count sort, and some models then passed sort: "citation", which failed…
- [App supervisor] The Environment and Proxy environment dropdowns in App Supervisor now sort names in ascending order, with numbered names such as cluster-2 listed before cluster-10.
- [ai] Fixed global MCP tool permission settings being lost after application restarts and becoming inconsistent across application instances.
- [UI templates] Fixed an issue where a duplicated popup could still show edits to the original page when its template was saved from an action that opened another popup.
- Captions on Windows. Generating captions no longer closes the app as it starts.
- Black preview with Chroma Key. On some Windows PCs the Chroma Key effect could leave every preview black and exports failing. That cause is fixed.
- When the graphics driver drops out mid-session, Concat notices straight away and tries again every two seconds, not on every frame.
- Give a bot a big job and let it split the work. Type /burst, or let the bot offer it: it previews the parts, runs them side by side and gives you one result.
- Ollama, found for you. MausBot finds Ollama running on your computer and adds it in one click, shows the models you've pulled, and lets you turn a model's thinking off or pick its level.
- Reasoning levels for OpenAI and OpenRouter API models.
- commit: 4a068d08cf9ea6633f6ee48cea800944bf70f746
- source tag: beta2026-11-11_20-33-19
- regions: cn global
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.3.3 NPM Tag: latest
- Add Piccolo 4B, 9B, and 27B as local decision models for tool review and context compaction, with BF16 and Q8_0 recommendations, runtime installation and updates, and Hugging Face or ModelScope…
- Browse, sort, and comment on QwenPaw Community posts from the Console, draft resource-linked issue reports with model assistance, and open the community editor.
- Add the optional OpenUI plugin to display interactive forms and components within replies, preserve their state during streaming, and submit actions to the originating conversation
- core: Keep retry backoff when a trigger fails right after reactivation (4be543b)
- d5fea53: Fix PostgreSQL 13 compatibility when persisting MCP server selections and sandbox info by using jsonb_set instead of JSONB subscript assignment.
- Nothing to do. The browser extension updates itself.
- The first chat searches after updating rebuild each chat's search index once, in the background. Results fill in as it goes.
- Scheduled. A new page in the sidebar lists the scheduled messages of all your chats, with when each runs next, how often, and its text.
- POST /api/v2/memory/delete
- POST /api/v2/memory/delete_by_ids: delete episodes by memory id, with
- Configuration: [memorize] delete_wait_seconds (default 10),
- 修复旧 profile 整行覆盖配置且缺少 ops-ptc 配置节时,批量调用引导被静默关闭的问题;ops-ptc 现从代码层默认开启引导,显式关闭、其它模式与遥测默认值保持不变。 @T-Auto
- 统一四平台验证矩阵与工具链配置,使用按任务隔离的依赖缓存,扩展跨平台 Rust 检查,并在 Linux 验证中先装载实际发布的 musl worker。 @Vickko
- Fix batching guidance silently turning off when an older profile replaces the whole row config without an ops-ptc section.
- B1–B3 now check local picture motion in overlapping windows and reject camera/full-scene wrappers and tiny long-span drift as substitutes for meaningful local animation.
- Muted ambience no longer counts as a continuous sound bed. Motivated cuts and premixed masters remain supported.
- B3 voice credits remain examples rather than provider requirements.
- Mac Intel
- analytics: Measure onboarding and first-week use. #664 (b3c7db9)
- Capture consented Nod onboarding profiles in Customer.io. #667 (776ddbe)
- Capture is the one-shot API again, which is the path that worked before 2026-11-05.
- The build refuses to make a second capture path. An SDK older than 14 now stops the build instead of
- The two things v0.14.1 got right are kept: a capture failure says it is not a permission problem, and
- 桌面客户端改为 Electron 壳,登录过期后自动续上
- 自定义 MCP 可填写 OAuth 应用凭证并完成授权
- 新增 octop migrate openclaw / octop migrate hermes,把原有助手迁到一个专家
- feat(credentials): set up keyless credentials from the v2 API, CLI and Sim MCP
- feat(credentials): keyless Snowflake, Microsoft app-only and Google keyless, GitHub and Bitbucket tokens, DocuSign JWT
- feat(mistral): extract PDF images as files and expose table format in Mistral Parser
- One place for every resource. The Resource Center now manages skills, connectors, and specialists — including Marketplace installation, also in the Web UI — with shared cards, filters, and detail…
- Scientific prompt optimization. Refine the current project's prompt from the composer with one click; review, edit, or undo the result, and save custom templates in Project settings.
- A stronger built-in Agent. Codex subscription accounts are now available for the main agent and subagents, Claude Fable 5.2 joins the model list, and long answers are no longer cut short at 4,096…
- Add Bach 1.0 Pro as a default Hypit dependency, with text, frame, and reference SVML surfaces and generation through HypiHub. Last-frame-only generation requires 4K.
- PixVerse V6 on HypiHub accepts seed and multi-clip. aspect-ratio="auto" requires at least one video reference, and multi-clip cannot be combined with video references.
- Add embeddable HTML document output and share HTML playback handling with Studio.
- No model-provider catalog or request-admission behavior changes are included
- Attest Linux Open Interpreter archives and their checksum manifest in the
- Verify that original build provenance before publishing existing artifacts.
- MCP with Chrome's remote debugging off: browser-use --cli-mcp now drives its own Chromium instead of stopping at the chrome://inspect steps, and tells your agent how to switch to your Chrome…
- Slow pages: CDP calls from the CLI and MCP wait 20 s, like the agent. A page frozen by a JS dialog fails fast with an error that says to handle the dialog.
- Windows: open() in MCP browser_exec code reads and writes UTF-8.
- Cost tracking and pricing: Fixed incorrect transcription cost for some FLAC files, which could push a key over its budget <sup>Also in 1.108.5</sup>
- Cost tracking and pricing: Changed per-second transcriptions to bill the usage.seconds the provider reports instead of the measured length <sup>Also in 1.108.5</sup>
- 资源 write:修正文件处理与父目录刷新的衔接,保留每个文件自身的摘要和索引工作;父目录独立刷新,不替代尚未完成的文件索引。wait=true 等待文件自身的处理完成,不等待父目录 L0/L1 和目录索引刷新,请求参数不变。
- 目录摘要:待刷新期间复用已有文件摘要;恢复 semantic.overview_sample_limit 默认 32 项的稳定采样,文件自身仍完整建立索引。#7380、
- 存储稳定性:修复 stat 按记录 ID 查询的兼容性、向量 Collection 生命周期,以及多写元数据不完整时的启动问题。#7377、#7355、
- CLI archives built from the Studio release studio-v2.60.0.
- Flashcards: generate recall cues from selected knowledge-graph propositions or a PDF/text file, or write cards by hand.
- Coding Questions: answer teaching questions and practice problems in a syntax-highlighted editor, with optional starter code, tutor-written tests, test output, and model solutions.
- Drawing Pad: work over the current question or practice problem with a pressure-sensitive pen, eraser, undo and a hand tool for panning. Your drawing accompanies the next answer or message.
- 80bac51: Move @arizeai/phoenix-otel to OpenTelemetry JS 3 (@opentelemetry/sdk-trace 3.x, @opentelemetry/api 1.10) so spans created by apps on OpenTelemetry JS 3 are recorded after register().
- Updated dependencies [80bac51]
- @arizeai/phoenix-client@8.0.0
- ci: simplify automerge authentication (5dca8b4c4) — Anwesh
- fix(work): report already-completed items and missing selectors in work complete (3bd3ac910) — Sarthak Chaudhari
- feat(triage): add bounded ClickHouse investigation evidence queries (9f7ef6887) — Anwesh
Nov 10, 2026
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- Support SubAgents(dynamic_agents=...) under durable execution
- Strip every default-ignorable character before ArtifactScanner phrase matching and flag hidden variation selectors
- Load one instruction file per directory and nested files by default in RepoContext, add walk_up, and have Coder build on them
- souls: a soul that fails to load throws at build
- api: a store option on generateText and generateObject (6236fb5)
- memory: report each turn's recalled memories and what the block shows (59dca59)
- OpenAI: Fixed keyless workload identity deployments answering 401 on image edits, audio, files, batches and other routes
- Other providers: Fixed ollama_chat/ streams giving every parallel tool call index 0, so clients ran only the first call
- Messages API: Fixed /v1/messages answering 400 for a param listed in the deployment's additional_drop_params
- Select subagent models by task in the release skill
- Report if a message from an agent steered the running turn
- Hide the driver's cursor tools from Computer Use agents
- briefing: say why a document's card is there (29f903d)
- briefing: show watch suggestions in Needs you (a370704)
- briefing: show what As me did, and what it is watching, on Home (5bcea3c)
- fix: backport 1.13.1 fixes and image security updates to 1.12.9
- fix: absorb lxml and Mako image workarounds for 1.12.9
- fix: authenticate starter project MCP clients in auto-login mode
- Agents act with their person's rights. My Kortix uses every connector and secret you can use and posts or edits a Slack message itself, without starting other agents.
- The connector catalog lists apps; it no longer decides what may connect. Wispr Flow connects through its own server in one click, ahead of Composio.
- Sessions start without a repository checkout; repository access comes from grants.
- feat: create a signing secret from the webhook dialog
- feat: let GitHub bots resolve PR review threads
- feat: record one model.call event and metrics per provider attempt
- Fix: install requires-python is capped to >=3.10,<3.15, so uv tool install graphifyy (and pip) on a fresh machine no longer selects Python 3.15 — for which tree-sitter has no wheel — and then…
- Fix: cpp a plain class/struct forward declaration folds into its unique exactly-qualified definition when one file reaches the other through unconditional exact quoted includes, and a…
- Fix: cpp a member call resolves through a scoped type alias (using Alias = ns::Type;) and through inherited fields/methods, and a forward-declared pointee type is completed at the use site.
- feat(cli): merge feature/cli-0.3.0 into main (b80085bcd)
- feat(models): give Model.summarize() a client-side base implementation (75a0cdd99)
- feat(vended-tools): make the shell tool's default timeout configurable (e0a6a229a)
- Full Changelog
- 🎙️ Dictate with less setup: Voice settings suggests an on-device model for your language and offers one setup button.
- 📱 Pick up where you left off: Chats show their saved recent messages after a phone relaunch while reconnecting.
- ⌨️ More control from your phone: The terminal key row gains Shift, Ctrl and Alt latches and an optional Smart rail with agent actions.
- Stop the run after finish_pr_review
- Stop false "ended without calling finish_scan" error on PR reviews
- feat(cli): add -r and a title column to the resume picker
- Changelog for v0.67.0-preview.0
- Changelog for v0.66.0
- docs(readme): clarify Gemini CLI scope, limitations, and tool landscape
- New sandbox backends for SandboxShellTool and SandboxCodeTool: E2BEnvironment (ag2.extensions.e2b) runs them in E2B cloud sandboxes, and ag2.extensions.boxd runs them on boxd Linux microVMs.
- MCP Apps work with browser-based hosts: /mcp no longer redirects POST and the CORS preflight, and the docs include a CORS recipe.
- Extended prompt-cache writes are reported separately: Usage.cache_creation_extended_input_tokens counts OpenAI 12h and Anthropic 1h writes, which are priced differently.
- Added shell completion scripts via dcode completion.
- Fixed detection of editor launchers such as code.cmd on Windows.
- Improved dcode completion output and setup steps, and fixed Bash completion for --flag=value arguments.
- Private inspection stays private: paired-node photos and screenshots are no longer automatically attached to replies, and explicit media-read denials are honored when preparing reply attachments.
- Responsive replies and pages: long zero-width Telegram table cells and malformed plugin SVG transforms no longer stall the Gateway or browser page.
- Linked accounts survive resets: sessions keep their person-linked model account through /new, /reset, and idle or daily rollover. Fixes #179432. Thanks @masatohoshino Authored
- LiteLLM usage adapter (@oh-my-pi/pi-ai/usage/litellm) reporting a gateway's key and user USD budgets
- UsageApi.replaceConfiguredProviders registers configuration-sourced usage providers between runtime overrides and the built-in resolver, and UsageProvider.usesRequestKey makes aggregate usage…
- Providers with a login-warning auth rule show their account risk before a new login and start it only after the user answers yes.
- Core: In Safari, a small video file keeps its preload, and a clip that is playing keeps its source (cda84222a, #7498).
- Studio: The Studio dev server now checks the Host header on its API routes too, as the preview server does since 0.9.30.
- Studio Server: Thumbnail links take their origin from the server, and previews from other sites open sandboxed (983e865ad, #7492).
- This SDK release bundles CLI version: 0.27.2
- Source: CLI built from source (same branch/ref as SDK)
- This SDK release bundles CLI version: 0.27.1
- 4311fa1: Use Zod Mini for internal validation schemas to reduce bundle size while preserving the public message schemas' classic Zod methods, error messages, and compatibility with supported Zod 3…
- d9dd2b7: keep the request options when the useChat queue sends a tool approval first, and hold the queue while an async sendAutomaticallyWhen decides
- Updated dependencies [4311fa1]
- Choose Simple or Advanced again. The switch is back in the app. If you were on Advanced before 0.1.151, you're put back on it.
- No more stopping at 5 bots on Starter. Create any number; 5 stay awake at a time and 2 work at once, as the plan page says.
- ModelGo is a new AI provider. Add a ModelGo API key in Settings and pick from the models your key offers. ModelGo sponsors MausBot, and you'll see a one-time card about it.
- evolveNumeric accepts an opt-in surrogate config. Each generation proposes several times more mutation candidates than it evaluates.
- The surrogate only orders candidates. Every archived score still comes from the external evaluator, the evaluator budget per parent is unchanged, and any surrogate failure falls back to the…
- createFannSurrogate() supplies a model backed by ruv-fann@0.1.2 (Rust compiled to WebAssembly). ruv-fann is an optional peer dependency that needs Node 20.9 or later.
- Antigravity no longer denies every tool call after an upgrade from 14.x: an in-place upgrade from an older install left a PreToolUse hook (and a PreInvocation context hook) in…
- Antigravity: a model call that only called tools no longer becomes an "AntigravityProvider returned Completed" memory.
- Redaction: with CLAUDE_MEM_REDACT_ENABLED=true, server capture now applies the configured redactor before the capture leaves your machine, and to session and event metadata before it is stored.
- @claude-flow/shared 3.0.8 on zod 4. Fixes the 4 high-severity npm audit findings (GHSA-mrhx-c7xc-m3w5) on a fresh ruflo install.
- Pinned-key verification for pattern downloads. With requireVerification on, downloaded patterns must verify against a pinned key. Off by default.
- Helper files are written without following symlinks. If .claude itself is a symlink, the helpers are skipped.
- Playwright 1.6x: The browser automation stack is upgraded to the Playwright 1.6x line. Nova Act now requires playwright>=1.60.0.
- Skip SSL Pre-Flight Validation: A new skip_ssl_precheck parameter on the NovaAct constructor lets you skip the SSL pre-flight check, useful when connecting through environments with custom…
- Typed Service-Unavailable Error: Transport-level failures now raise ActServiceUnavailableError, a subclass of ActServerError.
- Changes since langchain==1.4.9
- release(langchain): 1.4.10 fix(langchain): keep middleware-internal model calls off the messages stream chore(deps): patch newer-python security lock resolutions fix(langchain): omit trace inputs…
- Static provider credentials without placement metadata now resolve only in the Authorization header or through explicitly enabled body/WebSocket rewrite policies, so custom headers such as…
- REST endpoints now return 403 for HTTP upgrade requests in both enforce and audit modes. Configure WebSocket traffic with protocol: websocket and explicit handshake and message rules.
- OCSF sandbox attribution moved: supervisor events use device.container.uid and device.container.name, gateway API Activity and Detection Finding events use resources entries with type: "sandbox"…
- 5a714d180e9076a3f5f761c1fc9ad04e887ce1fa fix: improve IP address parsing
- 31a10f155760d3ff8b0442d3a587c164e9af0bc5 fix: improve RADIUS accounting handling
- 70b103a26f1d523d81470f26ead9a4eff3393c36 fix: improve SAML message decoding
- The automated-review note is now a small footer. Posted reviews and inline comments no longer open with a "Note" callout; one line of small text at the bottom says the review is automated.
- The doorman stops even while a dashboard is connected through it. A terminal or control WebSocket open through the doorman kept it from finishing its shutdown, so a stopped doorman could linger as…
- The server no longer leaks three open files for every terminal it starts on macOS.
- Pi 1.5.2.
- Resuming a session whose project folder was moved or deleted now offers Continue, Fork or Cancel instead of only continuing; Fork copies it into the current folder so later resumes no longer ask.
- The Anthropic default model is now claude-opus-5-5.
- Preserve proxy 407 errors for unsupported authentication challenges without waiting for unused target TLS configuration.
- Sandbox config: add userPolicy.filesystem.enumeratePaths (Windows only), granting enumerate-only access to the listed paths.
- Fixed: an empty sendMessages batch after a completed Claude turn no longer fails with "This model does not support assistant message prefill".
- Minimal deferred MCP tools in ToolExecutionOperation
- Configurable and per-start voice instructions for live-voice
- Configurable GPT-Live endpoint, model and key
- Add IPv4/IPv6 CIDR support for sandbox host allow and deny rules in settings.
- Add support for Claude Fable 5.2 in the model picker under New models
- copilot help input now says that Backspace at the start of shell input leaves shell mode and keeps the typed text, not only on an empty prompt
- Installed plugins can be turned off with the disabled_plugins setting: a disabled plugin stays installed but contributes nothing to sessions
- Plugins can be off by default (defaultEnabled: false in their ai.mistral.vibe extension) and turned on with the enabled_plugins setting
- /skills lists your installed skills, one row per name, with search, a Space toggle to turn each one on or off, and the highlighted skill's body under the list or beside it on wide terminals.
- Added swarms as a preview: start one with /swarm <task> and the agent becomes a coordinator that adds worker agents, each in its own conversation, which work on the task in parallel and…
- Added /swarm board, a full-screen live view of the current swarm that also opens with alt+s (rebindable as swarm.board), with a Roster of members showing each one's state and model, a Posts feed…
- Added a ● Swarm(<title>) row to the active-items tray below the prompt while a swarm runs, showing the latest post and the post count and then Done when the swarm finishes.
- fix(xml): honor positional tuple schemas
- fix(adapters): preserve nested XML enum and null guidance
- Default LM API-family selection to native lm15 routing
- [Access control] Document and test default-role fallback for the reserved union header.
- [resourcer] Resource API requests with malformed JSON in the filter query parameter now return HTTP 400 instead of an empty success response.
- [client] Configure popup mode and size for linked association tags.
- docs drop 0.9
- docs: reading progress bar under the navbar
- fix(admin): hindsight-admin cloud is listed in --help
- Mobile discards sign-ins that finish after a server change, preventing the previous server's session credentials from reaching the new server.
- Credentials passed in local MCP server launch arguments are hidden from tool output, errors and run history, including authorization flags, numeric secret values and passwords in URLs.
- Incus computers explicitly start as an unprivileged user, including custom images without a declared user.
- It can now carry something across turns. 「帮我在群里问一下老王那个报价」 used to end when the turn
- 群 B 的小王 and 群 C 的小王 are no longer the same person. A nickname carries the conversation it
- Merges can be taken back. Two records that turn out to be one person can be folded, and undone for
- feat(web): EmptyState with illustrated drawings
- feat(web): LangfuseIllustration with lazy product scenes
- feat(topics): filter traces in the Topics setup flow
- #4762 e1912c8 Thanks @sailist! - Make steering interrupt active model requests and continue in the same turn, and add POST /api/v1/sessions/{session_id}/steer for one-request message steering.
- #4808 e3c3a12 Thanks @liruifengv! - web: Improved interactions and fixed known bugs.
- #4804 c50a55b Thanks @liruifengv! - Fix stdio MCP servers configured with script commands such as npx failing to start on Windows.
- Raise the minimum supported openai SDK version from 2.6.0 to 3.26.0 to support the Decisions API.
- Declared pydantic>=2.12 (>=2.14 on Python 3.15) as the minimum supported version.
- If you don't set keep_default_na or na_values in read_excel_kwargs, XLSXToDocument now keeps cell text such as "NA", "N/A", "NULL", and "None" instead of treating it as missing.
- Fork sessions whose folder is gone: when a session's original folder was moved or deleted, you can now fork it into the current folder instead of only continuing, so later resumes no longer ask.
- When a session's working directory no longer exists, for example after the project folder was renamed, interactive mode now offers Continue, Fork, or Cancel instead of only continuing.
- Fixed /import of a session whose working directory no longer exists leaving a copy in the session directory, and continuing it leaving a second copy with the same ID (s.jsonl and s-1.jsonl).
- Mac Intel
- Cheaper background work. Background checks cost less, and Usage shows background spend on its own.
- Clearer model errors. Quota and rate-limit failures from your model provider now say what happened.
- [Marketplace] Browse skill catalogs, external skill sources, Agent templates, and MCP resources in one Marketplace, with guided installation and focused MCP and skill management.
- [Notifications] A new notification drawer beside the sidebar gathers task results and official announcements, with unread indicators, reply previews, and announcements shown in your language.
- [Scheduled Tasks] Scheduled tasks now have a dedicated sidebar page with task conversations, paginated run history, and management for completed or missed tasks.
- core: Limit resume attempts for waiting and queued executions (106a957)
- npm: bash npm install -g base44@latest
- Homebrew: bash brew install base44/tap/base44
- Version: 0.3.2 NPM Tag: latest
- GitHub Copilot Auto: Pick Auto in the model picker when your Copilot account has an Auto pool, and Copilot's router chooses the model for each session at the Auto discount.
- Side Panel Tab Shortcuts: Right-click a side panel tab and choose Close tabs to the right.
- Mark Sessions Read or Unread: Open the menu on a tab, Home row, project row or inbox card and choose Mark as unread to keep its dot until you next visit.
- e77e13abcacbab666af7a88238a587c73306c35c MUL-6907 feat(i18n): add complete Russian localization for web and desktop
- d6a9f0ec161b45b1172b97ce316c6395a1cf34e7 MUL-8043: feat(issues): include sub-issue runs in execution logs and timelines
- 632445ec4e36f3aec990846581dcc5ef46c27a21 MUL-8165: launch the new homepage, with one URL per language
- Add the approved 63.53-second Jensen V13 finished editorial reference, true full-frame references, portable implementation and closed file fingerprints.
- Require story-led physical action chains, contact/force/secondary motion, varied silhouettes and title-aware framing instead of repetitive poses beside popping props.
- Clarify image-generation vs SVG routing for the bottom-right presenter; anonymous code stickmen remain on the left. Preserve approved likeness, narration and unaffected layers during local repairs.
- MCP without Chrome: browser-use --cli-mcp on servers, containers and CI boxes now drives its own Chromium, installed on first use. When no browser can run, you get one clear error.
- http_get in the CLI and MCP always fetches from your machine.
- mcp with allowed_domains can no longer be bypassed through a cloud browser.
- commit: 999f7b8b415b17d246acaed7fe07880439fcaf72
- source tag: beta2026-11-10_15-36-19
- regions: cn global
- Deep Scan runs independent Standard reviews in batches. Workers can share leads and verify each other's evidence.
- External scanner imports are enabled by default in the CLI, SDK, and plugin. You can preview supported reports, save findings locally, and publish supported findings to Cloud.
- scans logs finds sessions faster in large compressed histories. The GitHub Action also includes scan scope limitations in its job summary, including when a scan returns zero findings.
- 外出连接更稳(#1862):设置 → 手机连接 → 外出连接新增「外出时保持电脑唤醒」开关,默认开启,连接和重连期间电脑不自动休眠(屏幕仍可关)。桌面服务重启会沿用原来的 cloudflared 隧道和地址,手机不用重新扫码;隧道地址真的变化时,同一 Wi-Fi 下的手机会自动换到新地址,iPhone 还会收到「电脑地址已更新」的推送。
- 手机 P2P 直连(#1844,#1859):外出连接页新增「允许手机直连(P2P)」开关和状态行。开启后手机与电脑建立直连,聊天页菜单显示「直连 / 局域网直连」而不是「经 Cloudflare」;此前打包版读不到直连模块、永远只能走隧道兜底,现在打包版也能直连。手机把会话切到后台或锁屏几秒再回来,会原地续会话、只补收增量,不再重新拉取整段历史。需要新版手机 App(0.1.4)。
- 回合运行中切换执行权限(#1851):任务正在跑时,输入框工具栏的盾牌选择器可以直接在 ask 和 skip permissions 之间切换并立即生效,之后的工具调用按新模式执行;已经显示的审批卡仍等待你作答,不会被自动批准。
- fix(docs): polish code rendering and keyboard navigation
- fix(next): align with route-handler framing in 16.4.5 and hash Bun patches into turbo tasks
- fix(trigger): keep each worker's egress gateway address out of the env sync
- 从这一版起,更新引擎不再清空 Linux 环境。 这次运行时又有改动,所以还是下整包;但更新时只替换 App 管理的引擎部分(DSH、Node、pnpm、桥接和它们的目录),用 apt 装的工具、默认位置的 npm 全局包、Python / pip 包、配置和工作区都会保留。从 1.0.14 升上来就生效。Node 版本变了时,第三方原生扩展仍可能要重新编译。
- 如果更新停下来了,先别重置。 App 确认不了现有环境和新引擎兼容、更新清单损坏,或者旧进程没停干净时,会停止更新并说明原因,原环境原样保留。按提示重试;还不行就导出诊断并开 issue,不要重置环境或卸载 App。
- 匿名使用统计多报两项:手机型号和厂商系统版本。 型号是系统给的代号(如 V2405A),系统版本如 OriginOS 5.0、HyperOS 2.0,用来看问题集中在哪些机型和系统上。不含序列号、Android ID 这类能认出某一台手机的信息,仍可在「隐私与分享」里关掉。
- feat(scan): gate PLAN_TASK on file size
- feat(session): inspect saved sessions by ID from any directory
- feat(rules): add Clojure and ClojureScript review rules
- feat: edit executable plugin configuration in automation help
- feat: display conversation budgets with semantic meter fills
- feat: add Paper theme to Agent Canvas
- 1e47750: Add px dataset-versions list <dataset-identifier> with cursor pagination, optional limits, and pretty, JSON, and raw output.
- CLI archives built from the Studio release studio-v2.59.0.
- Mika-v0.1.39-arm64.apk:近几年的安卓手机基本都选这个,更小。
- Mika-v0.1.39.apk:通用包,所有手机都能装。不确定就选它。
- 两个包功能完全一样,可以互相覆盖安装;App 里「关于 › 检查更新」会自动选合适的那个。
- Support unified document and block breadcrumbs
- Support rich document titles and preserve slashes in notebook names
- Improve document tree browsing performance and access to large directories
- wren: serve a project-bound agent and governed MCP through a shared context layer (c4d6ac3)
- wren: address deferred review findings from the serve-agent context layer (b7f972f)
- wren: keep result columns that share a name in JSON output (2f70e4a)
No release notes published.
- Users table: the admin Users list, offboarding record and sign-in links
- Move model route resolution into a shared module and route decision rules through it
- Run the browser tool on the session's model route
- fix(cloud): keep Codex ChatGPT sign-ins alive in cloud workspaces
- feat(models): add Claude Fable 5.2
- fix(marketing): stop CCing the submitter on contact form emails
- Added fast mode on Vertex in the global, us and eu regions, falling back to standard speed if Vertex refuses it
- packaging: keep PyTorch extras out of all
- pipeline: say whose request an OutcomeSnapshot is (a2bcbd1)
- providers: add Hermes agent wrap support (7877241)
- Added Claude Fable 5.2 and Mythos 5.2 with medium default effort
- Both models require Claude Code 2.1.321 or newer; Mythos 5.2 appears only when Claude Code advertises access
- Microsoft Store app: updates through the Store.
- Desktop app installed from the installer (.exe / .dmg): no change in this release; it updates with the next bundled release.
- CLI: hermes update, or re-run the installer one-liner.
- fix(security): publish main builds only from pushes to main, and lift oauthlib to 4.0.1 (1f817d29a) — Vaibhav Upreti
- Added saved flows (experimental): record a browser task once, then replay its requests.
- MCP tools now list only the shared arguments each tool uses
- Fixed session reuse and relaunch, now decided in one place in the daemon. A connection target is kept only after an attach succeeds, and security settings persist across daemon restarts
Nov 9, 2026
- Shell login and logout: use pi auth login <provider> and pi auth logout <provider> without starting a session, with API-key and no-browser sign-in options. See Credential commands.
- Idle prompt cache warming: eligible caches now stay warm for up to 30 minutes between runs when a refresh is expected to pay for itself. See cacheWarming.
- Anthropic message threads: requests send only new messages while the server continues the conversation, with full-request fallback when earlier context changes. See Anthropic compatibility.
- Keep forked subagents on their parent's prompt-cache prefix.
- Preserve symlink names in virtual filesystem results.
- Correctly mark read_file media and Nemotron nudges as harness notifications.
- LiteLLM lists only the models the configured key can call (its authenticated /v1/models), keeping the proxy's metadata for them.
- In Tern, a turn's long-press menu on a phone can edit and resend a prompt, retry the turn, open the rewind page on it, or fork the session from it into a new session
- A rewind from the transcript says how many turns it dropped and offers Undo on its toast in Tern, which puts the turns and your draft back
- Skills: Route a style outside the templates instead of listing them (37f1819f3, #7450).
- CLI: Reverse and templates report their own runs; commands say which skills are installed (448f88ff3, #7444).
- Skills: Reverse-any-video and style-sampling install with the core set (ad6f7858f, #7445).
- macOS (-sea assets): unsigned (Apple secrets not configured)
- Linux (-sea assets): SHA256SUMS-sea published with the release
- AWS Bedrock: Fixed 400s from GLM, GPT-6 Luna and Nova 2 on Bedrock when the conversation carries Claude's signed thinking
- AWS Bedrock: Changed /bedrock pass-through routes to enforce the model's guardrailConfig over any guardrail the caller sends
- Vertex AI and Gemini: Fixed Vertex AI Claude batches failing with 404 at create
- Fable 5.2 is in the model picker. Choose it for any Claude bot. It has a 1M-token context and is cheaper than Fable 5.1.
- Simple is now the only mode. Everything you use every day is in one place, with fewer settings to look through.
- A Vault page. See and manage what's saved in one place. Secret and sign-in cards offer Use a saved one, so you don't paste the same thing twice.
- phoenix-sqlean: bump bundled sqlean to 0.29.0 (a0a8b49)
- Focus areas in My Work: My Work now opens on an overview of your focus areas. Each area gets a card with bundled concept artwork, its goal and live request counts; open one to see its requests.
- claude-mem can update itself: the worker checks once a day for the release promoted in releases/promoted-release.json, and installs only a newer version that npm shows as published by the release…
- Startup context is the timeline alone: the SessionStart message is now the header lines followed by the timeline, ending on the newest entry.
- Agents can run Arena races. A new arena tool (and glimmervoid arena) lets a session start a race in its own project, list Arena races, read one race with its contestants, finishes, diff stats…
- Export and apply settings from the command line. glimmervoid settings export prints your config.json as JSON with every secret removed.
- Glimmervoid's own tokens leave config.json. The PostHog keys, the Telegram bot token and the GitHub read token now live in secrets/config-secrets.json beside the config, readable by the owner only.
- fix(chat): settle turns that quits and crashes cut off, and offer to resume them
- fix(desktop): keep agents' browser panes reachable past the hidden-webview cap
- feat(desktop): unload hidden browser panes by memory budget, keeping pages Chrome would keep
- feat(models): add Claude Fable 5.2
- feat(library): Moving AI Agent Prototypes Into Production: A Practical Guide
- feat(library): 10 Self Hosted AI Agent Use Cases for Engineering Teams
- bash pip install claude-agent-sdk==0.2.180
- #433 758b1a3 Thanks @riverisbuilding! - Bring the destination project forward when switching tabs or opening a project from the plus button on Windows.
- #433 758b1a3 Thanks @riverisbuilding! - Reference original files and folders on Windows network shares that refuse system links. The references work for previews and exports.
- #433 758b1a3 Thanks @riverisbuilding! - Refresh open timelines after changes outside the editor, retry failed reads, and keep newer edits when an older response arrives late.
- Updated dependencies [e0eb120]
- Updated dependencies [5111655]
- @ai-sdk/gateway@2.0.178
- Added Claude Fable 5.2 (claude-fable-5-2), now the default Fable model — 1M context, $8/$40 per Mtok with $0.20/Mtok cache reads
- Added --settings <file> to claude remote-control, which applies a settings file to every session it starts; if the file changes, no new session starts until you restart claude remote-control
- Fixed MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them
- DiscordChannel.items() yields one DiscordMessage per message in the channel and one DiscordThread per thread, active or archived, including forum posts. Both have read_text() for embedding
- Each run lists the channel, and only the messages and threads that changed are reprocessed; a thread's messages are fetched only when the thread reruns.
- Misconfigurations that Discord answers with an empty or text-less success (the Message Content intent off, no Read Message History) and lost access raise an error, so the channel keeps its outputs…
- Opt-in context relevance operation
- Context window report and viewer
- AIHubMix declarative provider
- vibe-acp notifies the client when the administrator changes the organization's Vibe Code configuration.
- Vim Normal and Insert modes in the prompt input, enabled with the vim_keybindings config option
- The prompt of the conversation section at the top of the screen stays pinned while you scroll, and a click jumps back to it
- #2116 702d639 - Send request context to MCP servers as _meta. Set requestMeta on chat({ mcp }) or tools(), or pass meta to callTool, readResource, getPrompt, and the MCP Apps call handler.
- fix(renderer): unblock Feishin media player initialization #2468
- fix(media): complete audio constructors and native media ownership #2479
- feat(webaudio): expose native scheduled audio sources #2517
- Mac Intel
- onboarding: Add About you and Join us steps. #596 (94fee2e)
- Show that Nod is working at every moment. #630 (78fa31c)
- build(nix): install OpenBot 0.65.0
- Fix "Delete history" of an agent chat
- Do not show a save error after a color change of a remote agent
- Fix: watch ignored directories (node_modules, .git, build output) are pruned from the Linux watch tree, so a large project no longer registers tens of thousands of wasted inotify watches and risks…
- Fix: cpp a member-call chain routed through class fields declared in other headers (client.actions->Execute(), with value ., raw-pointer -> and single-arg unique_ptr/shared_ptr receivers) resolves…
- Fix: markdown a code span naming an uppercase \-namespace class (\Foo\Bar, App\API) or a type-qualified method call (Type::member()) is kept as a symbol mention instead of being dropped, while a…
- Anonymous count of the applications AIHawk sends. Each application that ends downloads one of these files, named after how it went: submitted (sent and confirmed by the page), unconfirmed…
- This release is not a version of AIHawk: it only holds the counter files.
- Add sandbox.userPolicy.filesystem.enumeratePaths in settings.json for Windows enumerate-only directory grants
- Add /index commands and enable managed shared tgrep indexing by default for eligible repositories
- PDF attachments on Grok 4.7 and Grok 4.6 fall back to file references instead of failing the turn
- Notifications are on for every project, and a new Inbox page in the sidebar replaces the bell.
- Artifacts no longer need a feature flag. Artifact comments have been removed.
- The Inbox moves out of the sidebar into the workspace menu.
- Default writer model: SMART_LLM is now openai:gpt-6-sol (was openai:gpt-5.4). Set SMART_LLM to keep the previous model.
- LLM idle timeout: an LLM call that sends nothing for 300 seconds now fails with a clear error instead of hanging the run. Tune it with LLM_TIMEOUT (0 disables it).
- generate_report now raises when every LLM attempt fails, instead of returning an empty report.
- This SDK release bundles CLI version: 0.27.1
- Source: CLI built from source (same branch/ref as SDK)
- This SDK release bundles CLI version: 0.27.0
- ghcr.io/omnara-ai/omnara-web@sha256:339ddbe2a7f35fd82deb7d41ccf3aa3665c059f89f54047e1cf6acc75ae58bb8
- ghcr.io/omnara-ai/omnara-api@sha256:7ab2f1557230c310814b619bbf2b5242097cfbefe0764cc4bdbcf09ccda74f20
- ghcr.io/omnara-ai/omnara-worker@sha256:c77f606387aa8eaa14dc28c982a7dce062d0d0cd2a1f481a7ffe8c5b70893934
- homepage is repository points at with directory: packages/workspace-probe; bugs points at the metaharness issue tracker.
- Description rewritten to state the package's one job in 168 characters; 9 keywords, each tied to a real feature.
- README opens with the one job, supported environments, a quickstart whose expected output was captured from a clean install of the packed tarball, and one link to the next package.
- Fix #2308: LLM facade drops opts.op before providers — per-op provider switches (e.g. think
- fix: restore OpenClaw installation and automatic memory hooks
- feat(llm): add configurable per-model Redis GCRA rate limiting
- ⚡ Find past work without slowing down: Session History searches run off the app’s main thread, keep results visible while you type, and show your last list immediately after a restart while it…
- 📱 Your agents, together on your phone: A new Agents tab lists agents across your connected desktops. Mark them read or clear them from the list, with activity status supplied by the desktop.
- 💬 Keep your chat moving: Switching Codex accounts moves your open Codex chats to the account you choose.
- feat(topics): proper cost tracking for executions
- feat(external-media-storage): make external media storage an enterprise feature
- feat(web): use Langfuse by ClickHouse logo in the app shell and auth screens
- Stores above 64 MiB keep working: whole-scope reads page, transport cap raised to 64 MiB
- Session list filters startedAfter, startedBefore, endedAfter, endedBefore
- AGENTMEMORY_SUMMARY_TIMEOUT_MS for large sessions on slow or local LLMs
- Nothing to do. The browser extension updates itself.
- Telegram is optional and off until you link your own bot. See Settings → Phone notifications → Telegram.
- Steps for you. When ChatGPT needs something only you can do (a sign-in, a USB key, a command it must not run), it hands you a step.
- Restore SJolt and APIMart logos and the second B5 preview across all eight languages using immutable CDN image URLs. Verified on the actual GitHub page.
- api: a Claude model on OpenRouter gets JSON mode when its strict schema exceeds Anthropic's union limit (a0f8ee6)
- memory: a blank line in working memory no longer ends the section (68867d8)
- memory: a zero share turns its section off, undefined keeps the default, report the shares used, keep graph associations whole, cut by code point, skip items that do not fit, and count ages in…
- MCP 与实时上下文:注册 Streamable HTTP MCP Server,检索、调用工具,并把工具挂载为可读取的 Markdown 节点。提供 REST、CLI 和 SDK 入口;同步上游工具后可查看受影响的挂载。
- Agent 规则查询:多种 Agent 集成增加 Rule Search 入口,在执行操作前按需查找适用规则。升级插件或重连 MCP 客户端后生效;这不会自动增加每轮拦截。
- Watch 多节点运行:增加 MySQL Watch 存储和列表分页;Connector 授权撤销后自动暂停,重新授权时只恢复此前自动暂停的任务。#6756、
- You can organize the sidebar into collapsible sections: create sections, drag Bots and groups between them, or let the Main Bot organize the team for you.
- Creating or editing a Bot puts its runtime (DeepSeek Harness model or a local agent CLI) in one selector, and you can set the reasoning effort of a Codex CLI Bot.
- Memory layers and tidy: entries carry an id, an optional end date and a pin; MEMORY.md writes spill to topic files when full instead of failing.
- docs: use ./start.sh as the quick start
- chore(deps-dev): Bump @playwright/test from 1.55.1 to 1.64.1 in /e2e
- chore(deps): Bump actions/upload-artifact from 4 to 7
- core: Limit resume attempts for waiting and queued executions (c7aa6aa)
- @metaharness/darwin is now a regular dependency of @claude-flow/cli. Before, it was optional. ADR-526 amends ADR-150 rule 2 for darwin only.
- In-process calls: metaharness_bench --op verify and all four metaharness_gepa ops (genome, validate, render, analyze) now run inside the MCP server on a cached darwin module.
- ruflo-metaharness 0.5.0: bench.mjs verify imports darwin directly and never installs anything when a tool runs.
- No model-provider catalog or request-admission behavior changes are included
- Add one canonical file-backed managed Goal creation contract for an explicit
- Make creation IDs durable across response loss and restart, with atomic
- It installs on an older phone now. minSdk 30 → 27, so Android 8.1 and up take the package instead
- …but on 8.1–10 it cannot see the screen. That is not a bug and it is not hidden: the app names the
- 唤醒词 严格 / 标准 / 灵敏 / 很灵敏. If it keeps missing your wake phrase, or keeps waking on the TV,
- 更新前请注意:这一版的运行时有更新,所以还是下整包。 内置引擎仍是 DSH 0.2.2-rc.1,变的是我们给它打的补丁和桥接层。升级后 Linux 环境里只保留 /root,用 apt 装的软件和放在 /usr/local、/opt 等目录的文件需要重新装。对话和设置不受影响。
- 这一版没有在真机上验证过。 后台屏实时预览、多个后台任务、按截图坐标操作、本机文字识别、扫码配对和伙伴小屋的新功能都只经过自动测试和编译检查。遇到问题请开 issue 反馈。
- 新权限和新组件: App 新声明了相机权限,只在扫码配对电脑时用;不给权限也可以从相册识图或粘贴配对链接。手机操控内置了 Google ML Kit 中文文字识别,默认开启:识别在手机上完成,截图和文字不会发给 Google,但 Google 的 SDK 会收集机型、应用版本、耗时等使用统计。不想要可以在手机操控页或「隐私与分享」里关掉「本机文字识别(ML Kit)」,关掉后不会加载它。
- CLI archives built from the Studio release studio-v2.58.1.
- Document the apps_not_connected error on /kickoff and /status
- Run jobs on cloud backends. Kubernetes, Hugging Face Jobs, Modal, and Ray join SSH and Slurm: approve the target, command, resources, and outputs before submission, recover the job after a…
- Built-in Open-Science Agent. A standalone built-in agent runtime with provider-neutral turn execution and validated tool handling is now the default for new setups (previously Claude).
- Faster macOS updates. Updates download only the changed blocks of the app image, falling back to the full download when needed.
- bash curl -fsSL | bash -s -- --release --version 0.1.2026.11.9
- Start OpenSRE
- bash opensre
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