YZ·RADARyz.mikail.net
Daily edition · 15 November 2026

Claude Code sandbox permissions tightened; Meta AI gets Live AI and research

No changes to the TOP 10 today either. Rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are unchanged from November 14: Claude Fable 5.2 leads with 65 points, followed by Gemini 4 Carbon at 60, Claude Opus 5.5 at 59, and GPT-6 Astra at 57. GPT-6.1 Sol and Claude Sonnet 5.5 are tied at 55, Gemini 4 Hydrogen has 54, GPT-6.1 Terra has 51, and Gemini 4 Argon and Muse Wonder Full have 50 points. The watchlist is also unchanged; Kimi K3.1 Pro ranks 11th with 46 points and is the strongest open-weight model. LMArena’s live page was blocked by CAPTCHA; the latest accessible view is dated November 11, where Gemini 4 Carbon, Gemini 4 Argon, and Muse Wonder lead. Baidu’s ERNIE 5.5 Flash Preview still has no Artificial Analysis score.

Claude Code is once again the day’s busiest topic. The 47 items from version 2.1.325, released November 13, that were missing from yesterday’s archive were added today, and most change permission and security behavior. Even with the sandbox enabled, Bash commands now go through the normal permission flow by default; anyone who wants the previous auto-approval behavior must explicitly set `sandbox.autoAllowBashIfSandboxed: true`. Scheduled tasks and `/loop` no longer run skills that declare `allowed-tools` or hooks without approval; a skill that needs to run unattended must be added to `permissions.allow` as `Skill(name)`. Background sessions started with claude.ai sign-in can’t use or bill against the API key in the launching shell. Bypassing deny and ask rules with compound commands, allowing the sandbox to write to `.claude` settings in added folders, and symlinked worktree paths have also been blocked. The VS Code extension now supports Console sign-in without creating an API key; prompt suggestions now also appear in the integrated terminal and can be accepted with Tab. Anthropic also extended free Claude for Teachers access for verified K-12 teachers in the US through July 30, 2028.

Meta released version 299.0.0 of its Meta AI app for iOS on the evening of November 13. Release notes in the Turkey store list a new look, Muse Spark on AI glasses, a faster voice mode in 13 languages, Live AI, which provides instant answers about objects seen by the camera, product comparisons on the web and Facebook Marketplace, and Incognito chats that don’t keep a record. US notes also add daily briefings, scheduled tasks and reminders, email and calendar connections, and the ability to create research reports, presentations, quizzes, or workout plans in a few minutes. Meta says features will roll out gradually by country and language. There were no model or price changes at OpenAI. Dot agents now have persistent instructions separate from general custom instructions; they can be edited from the dot tab under Customize > Behavior on mobile. The Commerce Partner Playbook was also published for businesses, bringing together product feeds, agent access, payments, Sign in with ChatGPT, and WebMCP in one place. There are no new official records from Google. Reports that Google AI Plus subscribers lost the ability to select the Pro model in the Gemini app and AI Mode in Search on November 13 are currently based only on 9to5Google and Notebookcheck articles, so they are marked unverified.

DeepSeek clarified billing for its built-in web tools, which it made available through the API on November 13. Each search that returns results, each image, and each successfully fetched page is counted separately from token charges; Search and Fetch cost $0.75 per 1,000 successful requests. Empty or unsuccessful searches are free. At Moonshot, Kimi Code 2.6.1 and 2.6.2 fixed long sessions taking seconds to open and getting stuck with empty responses, and restored live subagent updates in the Swarm panel. In 2.7.0-canary, the `follow_up_behavior` setting, which determines whether Enter queues a message or adds it to the active turn, and `timeoutMs=0`, which removes the time limit, are being tested. Alibaba’s 14 records are from Qwen Code v0.27.3; in AUTO mode, `git commit --amend` is no longer blocked outright and instead requires human approval. Z.ai’s AutoClaw added natural-language scheduling for one-time and recurring tasks and announced a new weekly 100M token campaign for November 16–22. Xiaomi’s MiMo Code CLI and SDK reached 0.1.17 on npm; login and MCP hints now show the `mimo` command. Tencent LoopForge added the ability to choose and pin an execution engine in conversation, concurrent scheduled jobs per task, and GitHub agent tools for managing issues and PRs; teamai-cli also added team tool examples for Claude Code and Codex. There are no new records from xAI. A total of 119 new developments were added to the archive today.

Today’s TOP 10

The ten best unique models by intelligence index; each model is counted at its highest-scoring setting.

  1. 01

    Claude Fable 5.2

    Anthropic · Claude=
    65pts

    Released on November 9, it leads with 65 points. Available on all paid Claude plans; API ID `claude-fable-5-2`, $8 / $40, with thinking always set to adaptive. The Claude Meetings beta (Google Meet first) opened to Team and Enterprise plans on November 12.Released: 9 November 2026

  2. 02

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    Available in Gemini API and AI Studio as `gemini-4-carbon` (paid tier only; $2 / $10 through November 30, $4 / $20 starting December 1). Antigravity’s new Orchestrator mode assigns planning and difficult decisions to Carbon, and routine editing and tests to Hydrogen; Orchestrator (Carbon) is on Ultra.Released: 28 October 2026

  3. 03

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Anthropic’s flagship Opus model; available on Claude Pro/Max/Team/Enterprise and via the API. Claude Code 2.1.322 added fast mode support on Vertex (global, US, EU).Released: 22 September 2026

  4. 04

    GPT-6 Astra

    OpenAI · ChatGPT=
    57pts

    OpenAI’s most powerful GPT-6 model. The default reasoning effort for this model rose from `low` to `medium` in Agents SDK Python v0.24.0-alpha.7.Released: 3 September 2026

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

    The default model for paid ChatGPT plans; API pricing is $2 / $10. The default reasoning effort became `medium` in the Agents SDK alpha release.Released: 29 September 2026

  6. 06

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

    The most powerful model also available on claude.ai’s free plan; tied with GPT-6.1 Sol at 55 points.Released: 28 September 2026

  7. 07

    Gemini 4 Hydrogen

    Google · Gemini=
    54pts

    Released on November 12; scores 54 points with the xhigh setting and costs $3.16 per task ($5.48 for Carbon). Available as `gemini-4-hydrogen` in the API free tier too, with paid pricing of $1.50 / $7.50. Available through Firebase AI Logic and Android Studio. According to Google’s plan page, Hydrogen is also the top model available to Google AI Plus subscribers in the Gemini app.Released: 12 November 2026

  8. 08

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

    Released on October 26 in ChatGPT Work, Codex, and the API; focused on vision and computer use. The managerial preview of Dots agents runs through November 18.Released: 26 October 2026

  9. 09

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available on Google AI Pro/Ultra and the paid API; the $2 / $10 introductory price runs through November 30. Antigravity’s Orchestrator (Argon) is available to both Pro and Ultra subscribers.Released: 30 September 2026

  10. 10

    Muse Wonder Full

    Meta · Meta AI=
    50pts

    Meta’s most powerful Muse model; no public API. This week, the Muse app added a feature for paid subscribers in the US to call businesses, as well as the Muse from Meta Chrome extension.Released: 12 October 2026

Key signals

  • The TOP 10 is unchanged: Claude Fable 5.2 (65) leads, followed by Gemini 4 Carbon (60) and Claude Opus 5.5 (59). Kimi K3.1 Pro (46) ranks 11th.
  • Anthropic: In Claude Code 2.1.325, Bash commands inside the sandbox now require approval by default; skills with hooks in scheduled tasks need explicit permission.
  • Meta: Meta AI 299.0.0 for iOS adds Live AI, Incognito chat, voice mode in 13 languages, and research reports in minutes.
  • DeepSeek: Built-in web tools are charged per successful use; Search and Fetch cost $0.75 per 1,000 requests.
  • OpenAI: Dot agents now have separate persistent instructions; the Commerce Partner Playbook was published for businesses.
  • 119 new developments: Anthropic 48, Tencent 21, Alibaba 14, Moonshot 13, Meta 10, Xiaomi 6, Z.ai 3, OpenAI 2, DeepSeek 1, Google 1.

Worth trying today

01

Review your permission settings after Claude Code 2.1.325

Update with `claude update` and verify 2.1.325 with `claude --version`. If Bash commands inside the sandbox now require approval each time, that’s the new default behavior. If you deliberately want the old auto-approval behavior, add `"sandbox": {"autoAllowBashIfSandboxed": true}` to the settings file, but first check the sandbox’s write and network rules. If your scheduled tasks or `/loop` jobs use a skill with hooks or `allowed-tools`, add that skill individually to the `permissions.allow` list as `"Skill(skill-name)"`; don’t grant broad general permission. Before starting a background session, use `claude auth status` to confirm you’re signed into the right account. If you don’t want prompt suggestions, you can turn them off with `promptSuggestionEnabled: false`.

Anthropic · Claude
02

Keep DeepSeek web tool costs under control

Add the `deepseek_web_search` and `deepseek_read_url` tools to the `tools` list only for requests that need current information or URL reading; every successful search and fetched page is charged in addition to token costs (Search and Fetch cost $0.75 per 1,000 successful requests). Log the `tool_search`, `tool_fetch`, and `tool_image_search` counts under `usage.server_tool_use` in the response; note that the model can make multiple queries in a single call. Searches that return no results or fail aren’t charged.

DeepSeek · DeepSeek
03

Try Meta AI’s new features with care

Update Meta AI to 299.0.0 from the App Store on iPhone (iOS 18.4 or later is required). Open a new chat in Incognito mode for topics you don’t want in your persistent history. With Live AI, point the camera at a product or document and ask one clear question. For a research report, specify the topic, date range, and desired format up front. If you try connecting email and calendar, grant only the permissions you need. These features are rolling out gradually by country and language, so they may not be available in Turkey yet.

Meta · Meta AI
04

Give your ChatGPT dot persistent work instructions

If you have access to Dot, go to Customize > Behavior from the sidebar on mobile, enter your preferences in the task behavior field on the dot tab, and confirm. If that field isn’t visible, give the same instruction in a dot chat; for example: “When you finish the task, provide the result with a link and separately list anything that requires my decision.” These instructions are kept separate from general ChatGPT custom instructions. Set which actions require approval separately in Custom rules.

OpenAI · ChatGPT

This edition’s updates by AI

What changed in this edition for every AI in the TOP 10 and on the watchlist.

Anthropic · Claude

#1 Claude Fable 5.2#3 Claude Opus 5.5#6 Claude Sonnet 5.5Claude Haiku 5.5
Full history (836) →

New in this edition · 48

  • APIClaude Code

    Slash command permissions and feedback uploads are restricted for background agents

    A slash command’s `allowed-tools` permission is no longer accidentally applied to background agents woken during or after its turn. `/feedback` and `/bug` can now upload only sessions recorded by the signed-in account or organization.

    Usage details

    Who can use it: Claude Code 2.1.325; background agents, `/feedback`, and `/bug` flows.

    How to use it: If you start a background agent in a slash command, define each agent’s permissions separately; when submitting feedback, verify that the session was recorded under the same account or organization.

    Tip: On shared devices, check the upload preview before submitting a bug report; sessions belonging to another account are now blocked from being uploaded, but it is still best to keep the scope narrow.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Pricing / PlanClaude Code

    Background sessions are not billed to the wrong account or shell key

    Background sessions started with a claude.ai login can no longer use or bill the API key or token in the launching shell. Switching to another account in the background is rejected, the system explains what will run if no suitable credentials are available, and a second Enter in the exit prompt no longer accidentally stops the background task.

    Usage details

    Who can use it: Claude Code 2.1.325; background session, claude.ai login, and named profile flows.

    How to use it: Check the correct profile with `claude auth status` before starting a background session; if you need to switch identities, stop the session and explicitly log in again.

    Tip: For cost tracking, keep environment keys clean in the shell that starts the background service, and verify the organization billed for the session each time you start a new service.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Bash deny, ask, and allow checks are stricter for compound commands

    Fixes address some Bash commands bypassing deny/ask rules, compound commands running without approval, and allow rules running a different program from the one they specify. Commands using broad shell options and wrappers now request permission more consistently.

    Usage details

    Who can use it: Claude Code 2.1.325; Bash permission rules and shell wrappers.

    How to use it: Test rules with examples containing `sudo`, `env`, compound commands, and wildcards; if no permission prompt appears, inspect the command’s actual program instead of broadening the allow/deny list.

    Tip: Keep allow rules as narrow as possible, such as `Bash(program:*)`, and do not ignore startup warnings about variable wildcards.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    Bedrock provider restriction and credentials without HOME are fixed

    If `allowedProviders` does not permit Bedrock, token counting no longer sends the conversation to Bedrock. AWS `credential_process` and Google Cloud default credentials also work at startup when the `HOME` environment variable is absent.

    Usage details

    Who can use it: Claude Code 2.1.325; Amazon Bedrock, Google Cloud, and service environments without HOME.

    How to use it: Check the `allowedProviders` value in managed settings; when starting a service account, make AWS credential helpers or Google Cloud credentials accessible without depending on HOME.

    Tip: In CI, run a smoke test for the provider allowlist and a separate test of credential helpers without HOME.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Cloud connector approvals are no longer answered automatically while plugins load

    In cloud sessions, connector tool approval prompts are no longer answered automatically while plugins are still loading. Users can now see the actual connector and permission context after plugin loading finishes before making a decision.

    Usage details

    Who can use it: Claude Code 2.1.325; cloud sessions and connector tools that provide plugins.

    How to use it: If a prompt appears during a connector call, wait for the plugin loading indicator to finish, then check the displayed connector name and action before approving.

    Tip: Instead of calling plugins too early, use a health check at session startup to verify that loading is complete.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Sandbox cannot write to `.claude` settings and hooks in added directories

    As in the launch directory, the sandbox now rejects writes to `.claude` settings, skills, agents, commands, and hooks in added directories. This prevents a directory opened with `--add-dir` from silently changing session settings.

    Usage details

    Who can use it: Claude Code 2.1.325; macOS, Linux, and WSL sandbox sessions.

    How to use it: Add the extra directory with `--add-dir` and read its `.claude` contents if needed; if changes are required, make them in a trusted workflow outside the sandbox and with explicit approval.

    Tip: In shared repositories, put `.claude` changes through a separate review step; this is safer than granting write access across the entire working tree.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Account email is more strictly protected as a command and tool identity

    Account email checks now also reject its use as a client identity in commands run by scripts, file writes, web fetch operations, and MCP calls. This prevents the account’s email address from accidentally leaking into another tool’s identity field.

    Usage details

    Who can use it: Claude Code 2.1.325; script, file, WebFetch, and MCP calls.

    How to use it: Do not use your account email as a command argument or tool identity; define a separate, purpose-specific service identity for automations.

    Tip: In automations, replace your personal account email as a fixed identity with a short-lived service identity that has the fewest possible permissions.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Permission rules behave consistently across folders, plugins, and restarts

    After `cd`, relative deny/ask rules retain their scope in projects and added directories; differences in letter case in plugin MCP names no longer cause a rule to be missed. Deny rules are applied to approvals received after a restart, and `settings.local.json` committed in some subfolders is no longer mistakenly treated as a user setting.

    Usage details

    Who can use it: Claude Code 2.1.325; Bash, MCP plugin, self-hosted runner, and Remote Control sessions.

    How to use it: Specify an explicit path scope in permission rules and test `/permissions` behavior after a restart; use the same letter case as the actual server name in the manifest for plugin MCP names.

    Tip: Test relative rules such as `Read(./.env)` with added directories and worktrees; don’t assume they work only at the project root for security.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Plugin marketplace network connections and installation warnings are more secure

    When `ALL_PROXY` is used as the only proxy, plugin marketplace and plugin downloads now work through a CONNECT tunnel; if CONNECT is rejected, the operation fails explicitly. HTTPS authentication errors for private marketplaces explain that terminal prompts are disabled, and plugin installation warnings state that plugins can run code on the computer and access files and the shell.

    Usage details

    Who can use it: Claude Code 2.1.325; private plugin marketplaces and plugin installation.

    How to use it: If you use a proxy, define `HTTPS_PROXY` and add the marketplace again; review the manifest and source code before approving the code execution warning shown during plugin installation.

    Tip: For private marketplaces, use only a trusted Git source and an explicit `HTTPS_PROXY`; don’t mistake silent stderr output for a successful installation.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Bash commands inside the sandbox use the normal permission flow by default

    Even when the sandbox is enabled, Bash commands now go through the normal permission check by default, removing the assumption that commands will run automatically. The same behavior applies to managed settings, and `sandbox.autoAllowBashIfSandboxed` must be explicitly set to true to restore the previous automatic flow.

    Usage details

    Who can use it: Claude Code 2.1.325; terminal, Desktop, IDE, and managed sessions that use the sandbox.

    How to use it: If you intentionally want to keep running commands without permission prompts, define `sandbox.autoAllowBashIfSandboxed: true` in the highest-priority settings source; otherwise, approve commands as usual.

    Tip: In enterprise environments, audit write and external network rules separately before enabling this setting; the sandbox alone does not replace user approval.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    The sandbox can no longer write to updater and worktree settings files by mistake

    Sandboxed commands are now prevented from writing to the native updater download directory. Gaps in sandbox protection for worktree settings and hooks files, and for the linked worktree’s shared Git directory, have been closed.

    Usage details

    Who can use it: Claude Code 2.1.325; macOS/Linux/WSL sessions using the sandbox, worktrees, and linked worktrees.

    How to use it: Rerun updater and worktree operations on the latest version; if a write is denied, explicitly define the actual target path and the sandbox allowRead/allowWrite scope.

    Tip: When using a linked worktree, review the shared `.git` path as a separate security boundary; its presence under the project doesn’t make it safe.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    New security checks for worktree and self-hosted working paths

    Isolated worktree sessions reject `git bisect run`, `git maintenance register|start`, and Scalar installation commands, and explain the supported approach. Symlink hook paths in `EnterWorktree`, `.git` folders containing symlinks or hard links, and symlinked session/repository paths on self-hosted runners are now blocked as well.

    Usage details

    Who can use it: Claude Code 2.1.325; worktree, plugin hooks, self-hosted runners, and isolated sandbox flows.

    How to use it: Print the real path instead of using a symlink in a worktree hook; use a plain clone and avoid link- or hard-link-based layouts for `.git`. Run unsupported maintenance commands in the main trusted repository, outside the isolated session.

    Tip: Verifying self-hosted runner paths with `realpath` before starting catches paths that the new version will reject.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Managed plugin and skill restrictions now apply to sessions already in progress

    A bug that made the managed `strictPluginOnlyCustomization` restriction ineffective for agents and skills has been fixed. Even if the setting is applied after a session starts, the active session now enforces the organization’s rule allowing only approved plugin customization.

    Usage details

    Who can use it: Claude Code 2.1.325; organization sessions using managed settings, agents, and skills.

    How to use it: Set `strictPluginOnlyCustomization` in the managed settings.json and recheck plugin/skill behavior in an active session.

    Tip: After changing an enterprise setting, test existing sessions too; checking only when a new session starts isn’t enough.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Explicit permission is required to run skills with hooks in scheduled tasks

    Scheduled tasks and `/loop` no longer run skills that declare `allowed-tools` or hooks without user approval. To run a skill truly unattended, you must add `Skill(name)` to `permissions.allow`; the Always allow option in Desktop, IDE, and SDK is offered only if the persistent rule will allow the task to run again.

    Usage details

    Who can use it: Claude Code 2.1.325; scheduled tasks, `/loop`, Desktop, IDE, and SDK hosts.

    How to use it: Add `Skill(skill-name)` for the relevant skill to the `permissions.allow` section of your settings file, or approve the permission prompt when the task runs.

    Tip: In scheduled tasks, add only genuinely trusted skills to the allowlist by name; a broad automatic permission rule could run unexpected code.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    `claude purge` is more reliable without a terminal and with piped stdin

    `claude purge` no longer hangs silently when there’s no terminal to read from. The project picker and `-i` prompts also no longer ignore keyboard input when it arrives through an stdin pipe.

    Usage details

    Who can use it: Claude Code 2.1.325; headless and piped CLI runs.

    How to use it: For headless use, provide a TTY if interactive selection is required before running `claude purge`; explicitly route input for `-i` prompts in a piped stream.

    Tip: In CI, use a predetermined project scope instead of an interactive purge; remember that deletion is irreversible when automating deletion commands through a pipe.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    Subagent approvals, Agent SDK settings, and completed PR status are more accurate

    The main session is no longer reverted to its previous permission mode after a subagent’s plan is approved. In Agent SDK `applyFlagSettings()`, a single invalid value now produces an error instead of silently deleting previous settings; a completed pull request no longer makes the session appear to be unnecessarily waiting for input if it is only awaiting user review or a merge.

    Usage details

    Who can use it: Claude Code 2.1.325; Agent SDK, subagent plans, and pull request sessions.

    How to use it: Validate `applyFlagSettings()` calls in the SDK with valid values individually; after plan approval, check the active permission mode and what step the PR is actually waiting for.

    Tip: Adding cases for a single invalid field and a completed PR to SDK integration tests catches silent state changes.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AgentClaude Code

    Subagent model selection and truncated output behavior fixed

    Subagents using their own model are no longer switched to the wrong model by `opusplan` or `haiku` plan mode. When a subagent’s output is cut off at the token limit, it is now asked to produce a shorter, complete response instead of returning only the final fragment.

    Usage details

    Who can use it: Claude Code 2.1.325; subagents and sessions using the `opusplan` or `haiku` model setting.

    How to use it: Specify the desired `model` explicitly in the subagent definition. For tasks approaching the output limit, allow the subagent to produce a second, shorter response.

    Tip: Break long tasks into subagent assignments to reduce both the risk of truncated output and the cost of retries.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Background MCP tasks keep updating after connection loss

    Fixed an issue where background MCP tasks stopped updating when a later reconnection followed a connection loss caused by another call to the same server. A pending task question no longer appears over a message while the user is still typing; it waits until the draft is sent or cleared.

    Usage details

    Who can use it: Claude Code 2.1.325; background MCP tasks and terminal/host sessions running while a user draft is open.

    How to use it: Monitor the background MCP task from the `/tasks` screen; if you’re typing a draft, send or clear it first, then answer the question that appears.

    Tip: When a connection drops, waiting a few seconds and checking `/tasks` instead of canceling the task may preserve the result of the reconnected job.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude for Teachers

    Free access to Claude for Teachers extended through July 30, 2028

    Anthropic extended free access to its Claude for Teachers program for verified U.S. K-12 educators through July 30, 2028. The program continues to offer premium Claude capabilities, a Learning Commons connection, and teaching skills.

    Usage details

    Who can use it: Verified K-12 educators in the U.S.; Claude for Teachers individual teacher offer.

    How to use it: Go to Anthropic's Claude for Teachers page and verify your teacher account; once verified, use the free program access.

    Tip: Check school and district policies before using student data; Anthropic's page states that authorization is required for identifiable student data.

    ✓ Verified by official sourceSource: Introducing Claude for Teachers
  • FeatureClaude Code

    Code Mode worktree, approval, and resume flows are more controlled

    Code Mode scripts now correctly preserve the worktree selected with `EnterWorktree`, path resolution, and MCP session state. A bug that prevented the Yes text entered in the approval dialog from reaching Claude has been fixed; when a session that ends during a script is resumed, Claude no longer starts working on its own without a new prompt.

    Usage details

    Who can use it: Claude Code 2.1.325; experimental Code Mode and sessions using worktrees.

    How to use it: Enable Code Mode, enter a worktree, then run a script; after resuming, send a new prompt to continue and check MCP status with `/mcp`.

    Tip: Before using Code Mode in production automation, add separate regression tests for worktree paths, approval text, and waiting after resume.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Console login and browser verification no longer show false success after cancellation

    Pressing Esc immediately after the browser step during `/login` no longer accidentally changes the current login. Also fixed an issue where the browser displayed a success page for failed or canceled logins.

    Usage details

    Who can use it: Claude Code 2.1.325; `/login` and browser-based Console/Claude login.

    How to use it: Start `/login`, wait until browser verification is complete, and if you cancel, check the actual status in the terminal; if the wrong account was opened, switch back to the current profile.

    Tip: On shared machines, verify the correct account and organization name before closing the login window.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Desktop

    Desktop third-party deployments no longer get stuck waiting for token refresh

    Fixed an issue where sessions in third-party deployments of Claude Desktop remained in the `Stopping` state for a long time while waiting for token refresh. Sign-out and reconnect now complete faster.

    Usage details

    Who can use it: Claude Desktop third-party deployments that include Claude Code 2.1.325.

    How to use it: Update Desktop and restart the third-party gateway session; if the `Stopping` state persists, check the token refresh and gateway connection logs.

    Tip: Keeping the gateway token lifetime longer than the Desktop session duration reduces frequent shutdown and reauthentication loops.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Desktop

    Desktop and SDK sessions reconnect folder MCP servers correctly

    When a folder is moved or a session starts prewarmed, the folder's own `.mcp.json` and local MCP servers are no longer incorrectly listed as if they had been added by the host. WebSocket-based MCP servers are also connected correctly.

    Usage details

    Who can use it: Claude Code 2.1.325; Claude Desktop and Agent SDK sessions, moved folders, and WebSocket MCP.

    How to use it: After moving a folder, restart the session and check the `/mcp` list; define the WebSocket server in the project folder's MCP configuration rather than the host's.

    Tip: Checking the MCP list in prewarmed Desktop sessions before making an actual tool call helps catch an incorrect host/project distinction early.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Login startup and short auth commands run without locking up

    Valid keyless Console or Anthropic profile tokens can now be read without waiting for the credentials lock. Fixes also address a lock left by short commands such as `claude auth status` in a Console profile, as well as the accumulation of temporary OAuth lock files.

    Usage details

    Who can use it: Claude Code 2.1.325; keyless Console login, Anthropic profiles, and short CLI commands.

    How to use it: After updating, run `claude auth status`; if an old process is stuck, terminate it first and try the command again. For a new login, use `/login` or a Console profile.

    Tip: Running short auth checks before real requests during CI startup surfaces profile-lock-related delays early.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    Read-only Git operations in untrusted repositories now require approval

    Read-only Git commands in repositories that are not the project's own or that are untrusted now require permission first. Use `--add-dir` instead of `additionalDirectories` to run the command in a trusted additional directory.

    Usage details

    Who can use it: Claude Code 2.1.325; Git operations in terminal, Desktop, IDE, and cloud sessions.

    How to use it: Start Claude Code with `--add-dir /path/to/repository` to designate the repository as a trusted additional workspace; when prompted for permission for an individual read-only command, review and approve it.

    Tip: Keep the scope of `--add-dir` narrow instead of creating a persistent allow rule for untrusted repositories, even for read-only operations.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Reading your own Artifacts and cloud Artifact permission hooks work correctly

    Claude no longer asks for unnecessary permission when reading a user’s own Artifact or data. A fix also addresses cases where deny and ask rules and hooks for the Artifact tool were not applied in cloud sessions.

    Usage details

    Who can use it: Claude Code 2.1.325; Artifact tool and cloud sessions.

    How to use it: Open your own Artifact and run the tool call; to enforce an organization rule, define a deny or ask rule for the Artifact tool in the settings file and test it in a cloud session.

    Tip: Test the Artifact access policy separately with a real user and a cloud session; success in the local terminal alone is not enough.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Claude is informed when an MCP task is waiting for a response

    When an MCP task is waiting for a user response, Claude now knows the task is pending and that the `/tasks` screen can be checked. This reduces the chance that the model will assume a background MCP task is complete.

    Usage details

    Who can use it: Claude Code 2.1.325; terminal, SDK, and host sessions that use MCP tasks.

    How to use it: When an MCP task asks a question, open the `/tasks` screen and select the pending task; provide the required response to continue.

    Tip: For long-running MCP jobs, clearly specify up front in the prompt which steps require a user response; this makes pending tasks easier to track.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    MCP plugin reload and reauthentication states recover correctly

    Fixed issues where a removed MCP server remained in the `/mcp` list during plugin reload and a server continued to show needs authentication after a briefly rejected token refresh. SDK tool calls to a reconnected MCP server no longer incorrectly return `Not connected`; headless elicitation is also no longer silently canceled when the connection changes.

    Usage details

    Who can use it: Claude Code 2.1.325; plugin reload, token refresh, Agent SDK, and headless MCP sessions.

    How to use it: After reloading the plugin, refresh the list with `/mcp`; if authentication is required, use the Reconnect option and retry the SDK tool call.

    Tip: Don’t start many tool calls at once while an MCP server is reconnecting; verify the connection is ready with a single health check.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    Model identity, Fable limits, and provider authentication messages fixed

    A saved `sonnet[1m]` setting is no longer converted to a model identity rejected by the API on a new installation. `/fast` in a keyless Console session and `/model` on an unreachable network no longer show an incorrect organization restriction; the Fable limit message now explains the reset time and next steps.

    Usage details

    Who can use it: Claude Code 2.1.325; Sonnet 1M setting, Fable, and keyless Console/gateway sessions.

    How to use it: After a new installation, verify the model identity with `/model`; when you reach the Fable limit, follow the reset time in the message and check `/usage` or plan options.

    Tip: If you use a gateway, don’t mistake lack of network access for an organization policy error; first check the connection and the provider’s actual model list.

    ✓ Verified by official sourceModels: Sonnet 1MSource: Claude Code changelog
  • APIClaude Code

    OpenTelemetry prompt events include mid-turn and merged messages

    A `user_prompt` event is now emitted for prompts sent while tool calls are in progress, carrying `delivery.mode: mid_turn` and the active turn’s `prompt.id` fields. Also fixed an issue where earlier text was missing from telemetry events when queued messages were merged in VS Code or a message contained an image.

    Usage details

    Who can use it: Claude Code 2.1.325; Claude Code and VS Code sessions using OpenTelemetry.

    How to use it: Use a schema in your telemetry consumer that reads the `user_prompt` event’s `delivery.mode`, `prompt.id`, and merged text fields; update the old collector schema.

    Tip: Validate mid-turn and queue-merge scenarios as separate events in telemetry tests; don’t rely only on end-of-turn events.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Security review and inline Git operations work in special worktrees

    `/security-review` and inline Git commands in skills no longer stop with false errors in bare-repository worktrees, `--separate-git-dir` checkouts, and Android `repo` checkouts. Users can now continue working in these setups.

    Usage details

    Who can use it: Claude Code 2.1.325; bare repository, separate-git-dir, and Android repo checkouts.

    How to use it: Open the repository using the relevant checkout format and run `/security-review`; if the operation still stops, verify the displayed Git path and worktree status.

    Tip: Record the checkout type explicitly in CI before automating security reviews; this makes debugging much easier.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Sparse clone root files and adding repositories to self-hosted sessions fixed

    On older Git versions, sparse plugin or marketplace clones using `worktree.sparsePaths` no longer exclude repository root files. A repository added mid-session to a multi-repository self-hosted runner session is now available as well.

    Usage details

    Who can use it: Claude Code 2.1.325; older Git versions such as Git 2.34 and self-hosted runner sessions.

    How to use it: After updating the sparse configuration, clone again. In a self-hosted session, after adding a new repository, select it again using the repository picker or the relevant work command.

    Tip: On older Git versions, verify root files and plugin manifests in a CI step after a sparse clone.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Code

    Team sessions start in the correct folder and warn about temporary scratchpads

    New sessions for tmux and iTerm2 teammates, as well as sessions opened with the left arrow from the Agents view, now start in the session’s own folder instead of the shell’s last `cd` location. Claude’s instructions also say not to put files intended to be kept permanently in a temporary scratchpad folder that can be cleared.

    Usage details

    Who can use it: Claude Code 2.1.325; sessions using tmux, iTerm2, and the Agents view.

    How to use it: Open a new session from the Agents view or start a tmux/iTerm2 teammate; have persistent files written to the project folder or a designated plans folder.

    Tip: In automation prompts, specify the persistent destination path for outputs; this prevents work from being lost when the temporary working folder is cleared.

    ✓ Verified by official sourceSource: Claude Code changelog
  • APIClaude Code

    Context limits for unrecognized Claude models and mTLS password handling fixed

    `CLAUDE_CODE_MAX_CONTEXT_TOKENS` now also sets the context for unrecognized model IDs beginning with `claude-`. Claude Code's ripgrep, local Git, and sandbox helper processes no longer pass the mTLS client key password into their environments.

    Usage details

    Who can use it: Claude Code 2.1.325; unrecognized `claude-*` models used through a gateway and mTLS environments.

    How to use it: Set `CLAUDE_CODE_MAX_CONTEXT_TOKENS` to manage the gateway model's context; use `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` to retain the old 200K assumption.

    Tip: If you name gateway model IDs in the `claude-` format, verify the context value separately; the 1M default can cause overfilling with providers that have smaller windows.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Teleport safely stops stash selection in unusual Git states

    In some unusual Git states, `/teleport` now stops and tells the user what to do instead of attempting to select the wrong stash. This prevents working changes from being moved accidentally when resuming remotely.

    Usage details

    Who can use it: Claude Code 2.1.325; `/teleport` and complex Git states.

    How to use it: If `/teleport` stops, first inspect `git status` and the stash state, apply the recommended fix, then run the command again.

    Tip: Using a clean working state or a named stash before teleporting reduces ambiguous stash selections.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Code

    Prompt suggestions are more visible and less intrusive in VS Code and terminals

    Prompt suggestions are now also shown in VS Code and similar integrated terminals, where they can be accepted with Tab. When background or scheduled tasks start eight consecutive turns, suggestions refresh less often to reduce unnecessary usage.

    Usage details

    Who can use it: Claude Code 2.1.325; VS Code and integrated terminals that support prompt suggestions.

    How to use it: Press Tab when a suggestion appears in VS Code or the terminal. To turn suggestions off, set `promptSuggestionEnabled: false` in Claude Code settings.json.

    Tip: Keeping suggestions on and turning them off only for long-running automatic tasks preserves quick repetitions while reducing unnecessary refreshes.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    VS Code adds Console sign-in without API key creation and an MCP redirect field

    The Claude Code VS Code extension now supports signing in with a Console account that does not create an API key. In SSH, WSL, and container windows, there is a separate field for pasting the redirect URL when signing in to an MCP server from the panel. The untrusted folder notice explains why the extension is disabled and how to enable it; errors for blocked programs, localhost connections, models, effort, providers, and MCP permissions are also clearer.

    Usage details

    Who can use it: VS Code extension containing Claude Code 2.1.325; untrusted folders, SSH/WSL/container windows, and managed organization settings.

    How to use it: Open the Claude Code panel in VS Code; select a Console account at sign-in, paste the redirect URL in the MCP panel if needed, and approve trust for a new folder. While a command is running, check its output in the chat card and the model and effort options in their respective selectors.

    Tip: If your organization uses `allowedProviders` or `maxEffortLevel`, check your administrator's limits first; the extension now handles unavailable options more accurately.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Fixed Windows resume and Linux sandbox path edge cases

    Inaccessible PATH entries no longer cause unnecessary delays at startup on Windows; entries with different drive-letter casing or slash styles can be found with `/resume`, and project trust is preserved. On Linux, sandbox commands using the intersection of allowRead and denyRead under a symlink no longer produce a `bwrap` error.

    Usage details

    Who can use it: Claude Code 2.1.325; Windows `/resume` and Linux sandbox.

    How to use it: On Windows, use `/resume` to select the saved session again; on Linux, run the sandbox path under a symlinked home directory and check the access results.

    Tip: Normalize Windows CI paths to a single drive-letter format; on Linux, write the sandbox rule using the real path of the symlink target.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Responses no longer disappear in Agent view or narrow terminals

    Fixed scheduled task responses being hidden in Focus view, background subagent prompts not appearing when fullscreen was off, long options overlapping in narrow terminals, and first-run lines disappearing. Control characters in the welcome header and the red error shown for commands that only the user can run are now displayed more safely and clearly.

    Usage details

    Who can use it: Claude Code 2.1.325; terminal, Focus view, Agents view, and first-run screens.

    How to use it: Resize the session in a narrow terminal and open the prompt or Agents view; if a command can only be run by the user, type it into the prompt yourself to continue.

    Tip: Increasing terminal width before first run prevents both display overflow and shortcut lines from disappearing.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Code

    Atlassian, Asana, and Sentry connectors no longer incorrectly appear to fail in cloud sessions

    A working Atlassian, Asana, or Sentry connector is no longer reported as failed in a cloud session, even if the plugin lists it at a different URL. The mapping between the connector ID and the plugin URL has been fixed.

    Usage details

    Who can use it: Claude Code 2.1.325; Claude Code cloud sessions and related connector plugins.

    How to use it: Keep the plugin up to date and reload the connector in your cloud session; if the error persists, compare the actual connector URL with the URL in the plugin manifest.

    Tip: In custom plugins, keeping the connector URL consistent with the official connector ID prevents false error messages after updates.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Pricing / PlanClaude Code

    Spend errors for Cloud and Remote Control no longer show outdated administrator guidance

    The `ask your admin to raise it` guidance has been removed from spending limit and shared budget errors under Cloud sessions and Remote Control. Users now see a more direct error flow that explains their current limit status.

    Usage details

    Who can use it: Claude Code 2.1.325; Cloud sessions and Remote Control.

    How to use it: For a limit error, check `/usage-credits` or your organization’s usage page. If needed, submit a separate request to your administrator under the current organization spending policy.

    Tip: To resolve a budget error, check not only whether the limit can be increased, but also that you’re signed in with the correct account and provider.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Dataviz skill improves recommendations for part-to-whole charts

    The bundled dataviz skill's part-to-whole recommendations have been refined for static visuals and categories with a natural order, such as age ranges. This results in more suitable chart recommendations for the same data type.

    Usage details

    Who can use it: Claude Code 2.1.325; bundled dataviz skill.

    How to use it: When calling the dataviz skill, specify whether the data is part-to-whole or naturally ordered, and whether the output is a static visual.

    Tip: For ordered categories such as age bands, prioritize visuals that preserve the order over pie charts.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    Sign-in and Voice error messages identify which setting is blocking the action

    The `claude --desktop` sign-in error now identifies the API key, provider, or setting blocking account sign-in. `/voice` also clearly explains when a managed setting or project setting prevents changing modes. On macOS and Windows, users are also shown where to select a microphone when no microphone audio is detected.

    Usage details

    Who can use it: Claude Code 2.1.325; Desktop sign-in, `/voice`, macOS, and Windows.

    How to use it: Fix the blocking setting identified in the `/login` or `claude --desktop` error message. For Voice, select the correct microphone in your system sound settings, then run `/voice` again.

    Tip: First check whether enterprise settings override local settings; don’t try to bypass a managed rule by changing the local settings.json.

    ✓ Verified by official sourceSource: Claude Code changelog
  • CapabilityClaude Code

    First worktree run and brief sessions are more consistent

    The first `--worktree` run in a new home directory no longer skips the plugin's `WorktreeCreate` hook. Brief-mode sessions such as Dispatch also no longer send a second response after a turn response has been given.

    Usage details

    Who can use it: Claude Code 2.1.325; `--worktree`, plugin hooks, and Dispatch-like brief sessions.

    How to use it: In a new home directory, start `claude --worktree` with a plugin hook and verify the hook output; don't expect a second response for the same turn in brief sessions.

    Tip: Running the first-run test once with a clean home directory catches hook differences that appear in CI only with cached installations.

    ✓ Verified by official sourceSource: Claude Code changelog
  • AvailabilityClaude Code

    Slack app tips now account for cloud session policies

    Claude Code no longer displays Slack app tips when an organization policy denies cloud sessions. This avoids suggesting a workflow users can’t use.

    Usage details

    Who can use it: Claude Code 2.1.325; organizations that restrict cloud sessions and the Slack integration.

    How to use it: Check the cloud session permission in your organization’s settings. If it’s disabled, Slack tips not appearing is expected behavior.

    Tip: Verify your organization’s cloud session policy before covering the Slack workflow in user training.

    ✓ Verified by official sourceSource: Claude Code changelog
  • FeatureClaude Code

    Stats and Insights reports show more accurate information

    A bug that caused the Active days total in `/stats` to appear one day higher has been fixed. `/insights` no longer displays a report-ready message when there are no sessions suitable for analysis, and two values in the HTML report are now escaped safely.

    Usage details

    Who can use it: Claude Code 2.1.325; the `/stats` and `/insights` commands.

    How to use it: Run `/stats` or `/insights` in a session. If there’s nothing to analyze, build up enough sessions first, then try the command again.

    Tip: When exporting a usage report, interpret Active days and analysis eligibility as separate metrics.

    ✓ Verified by official sourceSource: Claude Code changelog
  • Plugin / IntegrationClaude Tag

    Claude Tag Slack Artifact links now show a card when sign-in has expired

    Fixed an issue where an Artifact link would not show a card if the poster's Claude sign-in had expired in a Slack workspace not managed by the organization. In this situation, the link now displays at least a plain-text card.

    Usage details

    Who can use it: Claude Code 2.1.325; Claude Tag, Slack, and Artifact links.

    How to use it: Open the Artifact link in Slack; if a rich preview doesn't appear, use the link in the plain card and renew your Claude Tag session.

    Tip: Don't rely on Slack previews; also include critical Artifact links directly in the message body.

    ✓ Verified by official sourceSource: Claude Code changelog
  • SettingClaude Code

    VS Code marketplace installation confirmation now uses a trust message

    The marketplace confirmation button in the VS Code installation link now says `Trust marketplace and continue` more explicitly. This change makes the decision to trust the extension source visible in the installation flow.

    Usage details

    Who can use it: Claude Code 2.1.325; VS Code extension installation.

    How to use it: When installing Claude Code from the marketplace, verify the publisher and select `Trust marketplace and continue` only if you trust it.

    Tip: Setting marketplace trust policies in advance on organization devices prevents users from accidentally approving the source one by one.

    ✓ Verified by official sourceSource: Claude Code changelog

Google · Gemini

#2 Gemini 4 Carbon#7 Gemini 4 Hydrogen#9 Gemini 4 Argon
Full history (244) →

New in this edition · 1

  • AvailabilityGemini uygulaması ve Google Search AI Mode

    Gemini Pro model selection removed from AI Plus on November 13

    Based on a notice sent to Google One AI Plus subscribers, 9to5Google and Notebookcheck report that as of November 13, 2026, Pro model selection in the Gemini app and Google Search AI Mode is available only on Google AI Pro and higher plans. Google’s current plan page shows access to Gemini 4 Hydrogen for AI Plus and expanded access to Gemini 3.1 Pro for AI Pro; however, because there is no official dated release note confirming the effective date, this entry is verified only by secondary news reports.

    Usage details

    Who can use it: Google AI Plus subscribers; the Gemini app and Google Search AI Mode. Sources report US plan prices; country availability has not been independently verified. Access to Gemini 4 Hydrogen and fast models remains available on AI Plus, while manual Pro model selection is reserved for AI Pro and higher plans.

    How to use it: Open Settings > Manage subscription in Gemini, or go to Google One > Manage membership. If Pro options do not appear in the model picker, upgrade to AI Pro in the Gemini app; also check the model options in AI Mode in Search.

    Tip: Gemini 4 Hydrogen or Flash access on AI Plus may be sufficient for short, fast tasks. If you need long-form research, complex coding, or specifically Gemini 3.1 Pro/Argon, first identify the tasks that genuinely require an upgrade.

OpenAI · ChatGPT

#4 GPT-6 Astra#5 GPT-6.1 Sol#8 GPT-6.1 Terra
Full history (339) →

New in this edition · 2

  • SettingChatGPT dot

    Separate custom instructions and Behavior controls arrive for ChatGPT dot

    You can set persistent instructions for how Dot should respond, work, or share status updates. These instructions are separate from general ChatGPT custom instructions; you can give an instruction in a Dot chat or edit it from the Dot tab under Customize > Behavior on mobile.

    Usage details

    Who can use it: Users with access to ChatGPT dot; on iOS and Android if Behavior is available on mobile, and on desktop and web through a Dot chat. Dots are rolling out to eligible users over 18 on Pro, supported regions on Business Premium, and eligible Enterprise, Edu, and Healthcare workspaces.

    How to use it: In a Dot chat, give a standing instruction, such as, ‘When you finish a task, summarize the result and tell me what needs my attention.’ On mobile, open sidebar > Customize > Behavior > the Dot tab > How [dot name] handles tasks, enter your preference, and tap the checkmark; if Behavior is not available, give the instruction in a Dot chat.

    Tip: Write instructions that specify scope, frequency, and expected output; define separately in Custom rules whether specific actions require approval.

  • Plugin / IntegrationCommerce Partner Playbook

    Commerce Partner Playbook brings together agentic commerce integrations for businesses using ChatGPT

    Alongside the ChatGPT Wallet announcement, OpenAI published an Agentic Commerce documentation suite covering product feeds, site access, payments, Sign in with ChatGPT, and plugin/WebMCP integrations. The guide helps businesses keep their products up to date in ChatGPT and make agent purchases work with their existing web and payment systems.

    Usage details

    Who can use it: Business owners, commerce developers, and websites that support product or payment flows in ChatGPT; this page does not specify separate plan or country restrictions.

    How to use it: Go to https://developers.openai.com/commerce and start with the Claim your business, Submit product feeds, Agent Access, or Enable browser access guides. For payments, implement Accept payments; for direct backend connections, Build a plugin; for flows that require an account, Sign in with ChatGPT; and WebMCP for supported transactions.

    Tip: First set up your business profile and an up-to-date price and inventory feed, then verify that ChatGPT agents can access your site. After that, add a checkout or plugin integration.

Meta · Meta AI

#10 Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 10

  • CapabilityMeta AI uygulaması

    Meta AI creates detailed research reports in minutes

    The 299.0.0 release notes list the ability to get detailed research reports in minutes. This makes it possible to start a multi-step information-gathering task with a single Meta AI prompt.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. The research feature may roll out gradually by country and language.

    How to use it: Ask Meta AI to create a detailed research report by specifying the topic, scope, and output format you want.

    Tip: Narrow the question using subtopics, a date range, and the types of sources you expect.

  • CapabilityMeta AI uygulaması

    Meta AI app brings the Muse Spark experience to AI glasses

    The 299.0.0 release notes say the app offers a smarter Meta AI experience with Muse Spark on AI glasses. This indicates that the assistant experience on AI glasses paired through the Meta AI app has been updated.

    Usage details

    Who can use it: AI glasses paired with the Meta AI app; iOS 18.4/iPadOS 18.4 and later. Features may roll out gradually by country, language, and device model.

    How to use it: Update the app, pair your AI glasses with Meta AI, and make a voice request to Meta AI through the glasses.

    Tip: Keeping both your glasses and the Meta AI app up to date increases the chance of getting the Muse Spark experience.

    ✓ Verified by official sourceModels: Muse SparkSource: Apple resmi lookup kaydı — Meta AI 299.0.0 (Türkiye)
  • CapabilityMeta AI uygulaması

    Ask about objects you see with Live AI in the Meta AI app

    The 299.0.0 release notes list Live AI, which provides real-time answers about what you’re looking at. This lets users ask about what they see as live context, rather than only uploading text or a photo.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Live AI may roll out gradually by country and language.

    How to use it: Update the app, start Live AI, point the camera at the object you’re interested in, and ask what you see.

    Tip: Keep your question focused on a single observation, such as the object’s purpose, how it differs from something else, or what to do next.

  • Plugin / IntegrationMeta AI uygulaması

    Meta AI can use email and calendar connections

    The Meta AI 299.0.0 release notes list connecting email and calendar accounts to help users stay organized. These connections make personal daily briefings and planning assistance more context-aware.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Connection options may vary by country, account, and supported services.

    How to use it: Update the app to 299.0.0 and grant the necessary permissions in Meta AI’s email and calendar connection flow.

    Tip: Grant only the account permissions you need and check the scope of the connected calendar or mailbox.

  • FeatureMeta AI uygulaması

    Meta AI offers daily briefings and scheduled tasks

    The official 299.0.0 release notes published for Apple’s US storefront list daily briefings, scheduled tasks, and reminders. This means Meta AI is no longer just a chat tool for instant answers; it can also help with recurring tasks.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Features may roll out gradually in select countries and languages.

    How to use it: Update to 299.0.0 from the App Store; ask Meta AI for a daily briefing or ask it to schedule a task or reminder for a specific time.

    Tip: Clearly state when you want the briefing and which topics it should cover.

  • CapabilityMeta AI uygulaması

    Meta AI voice mode now sounds more natural in 13 languages

    Meta AI version 299.0.0 release notes list a faster, more natural voice mode and support for 13 languages. This update expands in-app access for voice use instead of text.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. All 13 languages may not be available at the same time in every country or account.

    How to use it: Update the app, start a Meta AI chat, then turn on voice mode and select one of the supported languages.

    Tip: Using short, single-purpose voice commands improves response speed and conversational flow.

  • CapabilityMeta AI uygulaması

    Meta AI creates presentations, quizzes, and workout plans

    Apple’s official 299.0.0 release notes list Meta AI’s ability to create presentations, quizzes, workout plans, and similar outputs. This moves beyond chat responses to generating ready-to-use plans and content directly.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Features may roll out gradually in select countries and languages.

    How to use it: Ask Meta AI to create a presentation, quiz, or workout plan, specifying the target audience, duration, and format you want.

    Tip: Specifying the level, duration, output format, and evaluation criteria in your first prompt leads to more useful results.

  • FeatureMeta AI uygulaması

    Meta AI finds and compares products from the web and Facebook Marketplace

    Meta AI version 299.0.0 release notes list the ability to find and compare products across the web and Facebook Marketplace, and research purchases. This brings shopping search into Meta AI chat.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Product and Marketplace results may vary by country.

    How to use it: Type or say what product you’re looking for, your budget, and the features you want to compare to Meta AI.

    Tip: You’ll get more useful results if you specify your budget, delivery area, and must-have features upfront.

  • SettingMeta AI uygulaması

    Temporary, private Incognito Chat conversations arrive in Meta AI

    The 299.0.0 release notes list Incognito Chat for private, temporary conversations with Meta AI. This option provides a separate way to use the app outside the persistent chat history.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. The feature may roll out gradually by country and language.

    How to use it: Select Incognito Chat when starting a new chat, then begin your temporary conversation in that mode.

    Tip: Use this mode to keep sensitive or short-lived ideas out of your regular chat history.

  • FeatureMeta AI uygulaması

    New look released for the Meta AI iOS app

    The official 299.0.0 release notes published for Apple’s Turkey storefront say that the Meta AI app has a new look and new capabilities. This is a user interface refresh released alongside the specific features listed below for the same version.

    Usage details

    Who can use it: iPhone and iPad; iOS 18.4 or iPadOS 18.4 and later. Meta says some features may be limited or roll out gradually depending on country and language.

    How to use it: Update Meta AI to version 299.0.0 from the App Store, then reopen the app.

    Tip: Try features such as voice use, Live AI, and Incognito Chat from the main chat interface in the new look.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 13

  • RetiredKimi Code CLI

    /feedback now requires selecting diagnostic files and shows session files

    The text-only No attachment option has been removed from /feedback. Existing session files are previewed before submission; users choose either Current session logs or Session logs + project code.

    Usage details

    Who can use it: Kimi Code CLI 2.6.2; TUI feedback flow.

    How to use it: Enter /feedback; review the files in the preview and submit by selecting Current session logs or Session logs + project code.

    Tip: Use Current session logs if you don’t want to share sensitive source code.

  • CapabilityKimi Code CLI

    Failed or incomplete Kimi media uploads are no longer used

    Kimi Code now returns an error instead of caching a file reference when a media upload response is failed, processing, or invalid. Files that aren’t ready are not sent to the model.

    Usage details

    Who can use it: Kimi Code CLI 2.6.2; Kimi image and video uploads.

    How to use it: Attach an image or video; if you get an error before the upload completes, attach the file again and resubmit.

    Tip: For large media files, check that the upload has completed and the error details are clear before proceeding.

  • CapabilityKimi Code CLI

    Fixed long sessions getting stuck with empty replies on noted or errored tool results

    When tool results contain a note, error, or timing header, long sessions now continue compacting when needed instead of returning an empty reply.

    Usage details

    Who can use it: Kimi Code CLI 2.6.1 stable release and later 2.6.2 release.

    How to use it: Upgrade Kimi Code to 2.6.2; if needed, manually compact the session context with /compact and continue.

    Tip: If you see an empty reply during long tool workflows, check that you’re running version 2.6.2 before reopening the session.

  • Plugin / IntegrationKimi Code CLI

    WebSearch and feedback URL fields added for custom registry providers

    Custom registry providers in api.json can now route WebSearch and /feedback endpoints to their own services using the optional search_url and feedback_url fields.

    Usage details

    Who can use it: Kimi Code CLI 2.6.2; custom api.json registry providers.

    How to use it: Add search_url and/or feedback_url to the provider’s api.json definition; restart the CLI and call WebSearch or /feedback.

    Tip: Test the URLs alongside the provider’s env credentials and custom_headers settings.

  • AgentKimi Code CLI

    Live task updates from subagents return to the Swarm panel

    While Kimi Code Swarm is running, task events for subagents that start, continue, and finish are now sent to the panel, making live subagent status visible.

    Usage details

    Who can use it: Kimi Code CLI 2.6.2; sessions using the Swarm panel.

    How to use it: Enable Swarm mode, run parallel tasks, and keep the Swarm panel open during execution.

    Tip: For parallel work, use the panel to track subagents as they work and complete tasks, not just the final result.

  • APIKimi Code CLI

    timeoutMs=0 removes the time limit for subagents

    In Kimi Code 2.7.0-canary, subagent deadline calculation is disabled when timeoutMs is 0, allowing the subagent to run without a time limit.

    Usage details

    Who can use it: Only in the 2.7.0-canary.20261113055733-391d9e7 prerelease and in clients that programmatically use subagent calls.

    How to use it: Set the timeoutMs field of the Agent call that starts the subagent to 0.

    Tip: Use unlimited time only for work that genuinely takes a long time, and provide a separate way to stop stuck subagents.

  • CapabilityKimi Code CLI

    Token counting and automatic compact are now based on actual usage

    2.7.0-canary bases token counting and compact decisions on reported usage values, improving context management in long turns.

    Usage details

    Who can use it: Only in the 2.7.0-canary.20261113055733-391d9e7 prerelease; long-context TUI sessions.

    How to use it: Install the Canary version and run a long session; observe automatic compact behavior when the context threshold is reached.

    Tip: Track the context indicator during long tasks and keep critical instructions concise before automatic compact starts.

  • CapabilityKimi Code CLI

    TUI transcript history is preserved during trimming, folding, and replay

    2.7.0-canary prevents transcript history from being lost or corrupted during trimming, folding, and replay.

    Usage details

    Who can use it: Only in the 2.7.0-canary.20261113055733-391d9e7 prerelease; TUI sessions.

    How to use it: Install the Canary version and use /compact, transcript folding, and session replay in a long TUI session.

    Tip: If the history view changes during long sessions, first reopen the transcript and check the replay result.

  • SettingKimi Code CLI

    In the TUI, you can choose whether Enter queues or steers

    The new follow_up_behavior setting lets you choose whether Enter queues a message while Agent is running or steers the current draft into the active turn. The default queue behavior remains unchanged; the steer option injects only the current message.

    Usage details

    Who can use it: Only in the 2.7.0-canary.20261113055733-391d9e7 prerelease; Kimi Code TUI.

    How to use it: Go to /settings → Editor → Follow-up behavior and select Queue or Steer; alternatively, add follow_up_behavior = "steer" to ~/.kimi-code/tui.toml.

    Tip: Use Steer to interrupt a flow immediately, and Queue for work where you want to preserve message order.

  • CapabilityKimi Code CLI

    Fixed multi-second delays when reopening long sessions

    Kimi Code 2.6.1 fixed the multi-second wait when reopening long sessions; 2.6.2 also patched this path for long sessions with many tool calls.

    Usage details

    Who can use it: Kimi Code CLI 2.6.1 and 2.6.2 stable releases; upgrading to 2.6.2 includes both fixes.

    How to use it: Upgrade with npm install -g @moonshot-ai/kimi-code@2.6.2; then reopen the long session from /sessions.

    Tip: Upgrade the CLI to 2.6.2 before evaluating the reopening experience for large sessions.

  • APIKimi Code CLI

    The session’s request_id is added to /feedback submissions

    When submitting feedback, Kimi Code adds the session’s request_id to the diagnostic metadata, allowing support to link the review to the relevant request.

    Usage details

    Who can use it: Kimi Code CLI 2.6.2 and sessions using /feedback.

    How to use it: Complete the /feedback flow as usual; request_id is automatically added to the feedback submission metadata.

    Tip: Submit /feedback before reproducing the same issue to preserve the diagnostic context for the relevant request.

  • CapabilityKimi Code CLI

    Raw bytes are preserved in the final sections of task output

    2.7.0-canary preserves raw bytes when converting the tail of task output held in the queue to text, reducing the risk of corruption to special or non-text bytes.

    Usage details

    Who can use it: Only in the 2.7.0-canary.20261113055733-391d9e7 prerelease; task and background output.

    How to use it: Run a task or background command in the Canary version; inspect the final section of long output in the TUI.

    Tip: For tools that produce non-text output, save the final output chunk separately for debugging.

  • CapabilityKimi Code CLI

    TUI compact summaries keep Cron and MCP authorization calls separate

    Kimi Code 2.7.0-canary does not fold Cron and MCP authorization calls into compact transcript groups; these calls remain separate during summarization.

    Usage details

    Who can use it: Only in the @moonshot-ai/kimi-code@2.7.0-canary.20261113055733-391d9e7 prerelease; inclusion in the stable channel has not been confirmed.

    How to use it: Install the Canary package and run /compact in the TUI; check that Cron or MCP authorization calls remain separate from compact groups.

    Tip: When auditing automation and authorization traces, check these calls separately after compacting.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 21

  • Featureteamai-cli

    teamai status and push offer edited documents from docs mirror

    If a document the team already owns is edited in docs mirror, teamai status and teamai push now offer that edit at docs/<rel>. Mirror copies that match an existing or older version in the team history are skipped; edits in the checkout retain priority for the same path.

    Usage details

    Who can use it: teamai-cli teams with sharing.docs.localDir configured; new documents are added from the team repository’s docs/ directory, not from the mirror.

    How to use it: Edit a delivered document under sharing.docs.localDir, then run teamai status and teamai push; for a single file, you can use teamai push --file <mirror-file> --path docs/<rel>.

    Tip: Run teamai pull first to prevent a stale mirror copy from overwriting the team’s newer version.

  • Capabilityteamai-cli MCP

    teamai restores the ownership record when an MCP manifest fails

    If writing a project-scoped MCP manifest fails, teamai now attempts to restore its own ownership manifest to its previous state. If restoration also fails, it clearly reports which ownership records need repair, so incomplete install, update, or removal records are not silently left behind.

    Usage details

    Who can use it: Project-scoped MCP install, update, and remove operations.

    How to use it: Run teamai mcp install, teamai mcp update, or teamai mcp remove; if you get a write error, repair the reported ownership records and retry the same operation.

    Tip: Check write permissions in directories containing MCP files and check for concurrent worktree use in advance.

  • Plugin / Integrationteamai-cli

    teamai-cli adds team tool instances for Claude Code and Codex

    A team or member can now derive a tool instance from a published Claude Code or Codex definition. The instance can have its own command, home, and paths values, and can use separate session/key delivery. A command provided by the team requires approval on first use, and an unapproved instance is rejected when running unattended.

    Usage details

    Who can use it: teamai-cli team and personal scopes; instances can extend only Claude Code (claude) or Codex (codex) definitions.

    How to use it: Add schemaVersion: v1alpha1, name, extends: claude, command, and home fields to tools/acme-claude.yaml; then run teamai init --agent claude,acme-claude, teamai pull, and workflow start review --tools default=acme-claude.

    Tip: Provide command as a bare executable name on PATH, and review the YAML and executable before the initial approval.

  • CapabilityLoopForge

    Slow heartbeats no longer interrupt live LoopForge generation

    A late-arriving heartbeat no longer incorrectly marks an assistant generation that is still running as interrupted; the live turn’s workspace lease is preserved. When a dead turn is actually cleaned up, only its own lease is released, and MySQL checks are performed against the database clock.

    Usage details

    Who can use it: LoopForge conversation assistant generations under multi-replica or slow-heartbeat conditions.

    How to use it: In the LoopForge update, verify that migrations/032_assistant_generation_workspace_lease.sql has been applied; installations using AutoMigrate may add the schema at startup.

    Tip: Apply the migration before a rolling upgrade and test lease behavior with a long-running conversation generation.

  • AvailabilityLoopForge bind-cli

    bind-cli --service now also runs in the background in environments without systemd

    Even if the systemctl command is available, bind-cli --service now starts the runner as a background process instead of failing when the container has no actual init system. This makes it possible to bind runners in containers and similar environments without systemd.

    Usage details

    Who can use it: LoopForge runner binding; containers and servers without a systemd/init system.

    How to use it: Run deploy/bind-cli.sh --service during runner setup; if systemd is unavailable, verify the background process started by the script through the runner status.

    Tip: If persistence across container restarts is required, define your own supervisor or restart policy as well.

  • Capabilityteamai-cli

    Concurrent workflow claims preserve existing progress

    If two members or runners update the same workflow run at the same time during claim/recovery, newer local progress is no longer overwritten by an older snapshot. If progress has already been received, a second recovery preserves it; an actual state conflict is explicitly rejected.

    Usage details

    Who can use it: teamai workflow handoff, claim, recovery, and cross-runner publish flows.

    How to use it: For a delegated run, use teamai workflow resume <run> or the teamai workflow next flow; in concurrent claim situations, progress is preserved and there is no unnecessary restart.

    Tip: Instead of forcing the same run to execute on two runners, first check the current ownership and state conflict notice.

  • SettingLoopForge

    LoopForge locks the CLI selection after a conversation starts

    The CLI used by a conversation cannot be changed while a response is being generated or after the first response has been saved. To use another CLI, start a new conversation; the CLI locked to the conversation is shown in its title.

    Usage details

    Who can use it: LoopForge conversations with multiple executors/CLIs.

    How to use it: Select the executor from the CLI selector in the composer before the first message; after the conversation starts, create a new chat to use another CLI.

    Tip: For comparative testing, start a separate new conversation for each CLI to avoid mixing historical context.

  • APILoopForge API

    LoopForge multi-turn chat request limit raised to 32 MiB

    When tool activity in deep-thinking turns no longer fits within the 2 MiB default, an explicit 32 MiB limit can now be used for multi-turn chat, generation, and session bodies. If the caller has a stricter limit, it is not raised.

    Usage details

    Who can use it: LoopForge conversation/chat API clients; the gateway or client may impose a lower limit separately.

    How to use it: Send the chat, generation, or session request normally through the existing API; when using a supported body of up to 32 MiB, also raise the reverse proxy body limit.

    Tip: For large tool outputs, check client and reverse proxy limits together; trim unnecessary attachments.

  • CapabilityLoopForge

    LoopForge preview proxy restricts methods with an allowlist

    The development server preview proxy no longer forwards unsupported HTTP methods and rejects request paths containing line breaks. This gives preview routing a narrower and more predictable request surface.

    Usage details

    Who can use it: Workspaces using LoopForge side browser and dev-server preview proxy.

    How to use it: Open the preview through the usual side browser and use standard HTTP methods supported by the application; unsupported methods are rejected by the proxy.

    Tip: Explicitly allowlist methods in the preview integration and do not concatenate user input into a URL path.

  • Featureteamai-cli

    teamai status and doctor report docs mirror changes separately

    If a delivered document is changed in the docs mirror, teamai status can report it separately from normal checkout changes. teamai doctor also clearly identifies the mirror file and explains how to propose the document with teamai push --file.

    Usage details

    Who can use it: teamai-cli team documentation using sharing.docs.localDir.

    How to use it: Run teamai status or teamai doctor; propose the reported mirror file with teamai push --file <local-file> --path <team-path>.

    Tip: Don’t edit the checkout and mirror copies of the same document at the same time; keep clear which copy is the source.

  • Capabilityteamai-cli

    Recall derives code relationships from unquoted rule globs

    Recall relationships now use the same teamRuleData path as the rule renderer and include unquoted globs in YAML frontmatter in relationship calculations. For example, paths: **/shop/** can now generate a link to the relevant code page.

    Usage details

    Who can use it: Projects using teamai-cli recall and the source relationship graph.

    How to use it: Add paths: **/shop/** or a similar glob to the rule frontmatter and run the teamai recall flow; relationships are derived automatically.

    Tip: It’s still good practice to quote glob values to improve compatibility across YAML tools.

  • FeatureLoopForge

    Select and pin a LoopForge execution machine for a conversation

    You can select LoopForge Cloud or a connected execution machine before opening a conversation workspace. Once the first message is sent, the target machine is attached to the conversation and remains fixed for subsequent turns; the picker also offers actions to refresh the list and manage machines.

    Usage details

    Who can use it: LoopForge conversations; installations with a configured cloud pool or a connected online execution machine.

    How to use it: In a new conversation, open the execution machine picker at the top, select LoopForge Cloud or a named machine, and send the first message; the target is locked in from that point on.

    Tip: For long-running tasks, select an online machine with sufficient resources before sending the first message.

  • Plugin / IntegrationLoopForge

    LoopForge agent GitHub tools expand issue and PR operations

    Agent tools have been expanded to cover issue get/create and issue close, as well as viewing PR files and performing review, merge, and close operations. These capabilities also cover conversation/SOP agent flows, separately from the lightweight Main Bot tools that run without opening a workspace.

    Usage details

    Who can use it: LoopForge conversation and SOP agents with a connected GitHub integration.

    How to use it: Connect the GitHub integration and explicitly ask the agent to perform an issue or PR operation in the conversation; complete the permission/approval flow for write operations such as merge and close.

    Tip: Before asking for a merge or close, ask the agent to summarize the PR files and review status.

  • CapabilityLoopForge

    LoopForge scheduled AI jobs run concurrently per task

    Scheduled jobs can now run concurrently across different tasks instead of sharing a single workspace across the platform. The lock that prevents conflicts is now held per task rather than per user.

    Usage details

    Who can use it: LoopForge scheduled AI tasks and unattended runs.

    How to use it: Create and enable scheduled tasks through the existing interface or API; the scheduler manages concurrent jobs and locks for the same task automatically.

    Tip: Schedule independent tasks separately to reduce total wait time.

  • CapabilityLoopForge

    Native workspace file and command operations no longer require individual approval

    Native file, patch, and shell operations in the Docker workspace now run without waiting for approval for each operation. Actions that write to the platform or external tools remain subject to existing permission modes.

    Usage details

    Who can use it: Native workspace execution in LoopForge conversation workspaces; permission mode rules remain in place for writes to the platform and external tools.

    How to use it: Select the connected execution machine/native workspace target in the conversation; workspace file and command actions run directly, while external operations use the permission mode in Permission Picker.

    Tip: Keep writes to external systems in a stricter permission mode when using automated workspace operations.

  • CapabilityLoopForge

    A failed native session continuation is recovered once with a clean session

    If the journal is incomplete or the provider returns a continuation error such as 400/422, and the new native session has not yet produced output or a tool call, LoopForge starts a clean session once from the platform transcript. Identity, quota, network, or machine access errors instead preserve the session and report the cause.

    Usage details

    Who can use it: LoopForge native agent CLI conversations and session continuations after handoff.

    How to use it: Continue or resume the native conversation as usual; the clean-session fallback is applied automatically for eligible journal/provider continuation errors.

    Tip: Having the workspace state verified in the first response after a handoff helps prevent re-execution.

  • AvailabilityLoopForge

    Shared generic-docker sandbox added to the conversation workspace

    A generic-docker sandbox is now provided as a shared catalog profile for conversation workspaces. This reduces the need to create a separate sandbox profile for every project just to open a conversation.

    Usage details

    Who can use it: Installations using conversation workspace on the LoopForge platform; the change is applied through migration 030.

    How to use it: After updating LoopForge, verify that the migration has run and open a new conversation workspace; the generic-docker profile in the catalog can be used automatically.

    Tip: If you need a custom toolchain, define a custom Dockerfile sandbox profile instead of using generic-docker.

  • APILoopForge API

    LoopForge API asset endpoints switch to type-specific schemas

    The Rule, Skill, MCPServer, Dockerfile, Environment, and SandboxProfile save endpoints now define type-specific OpenAPI request body schemas instead of a generic Asset body. This lets clients validate required fields and asset types earlier.

    Usage details

    Who can use it: LoopForge OpenAPI clients and integrations that save these six asset types.

    How to use it: Refresh the OpenAPI definition and send the required fields for the relevant type when calling upsertRule, upsertSkill, upsertMcpServer, upsertDockerfile, upsertEnvironment, or upsertSandboxProfile.

    Tip: Regenerate automatic client code from the new OpenAPI schema; don’t blindly send the generic Asset model.

  • APIteamai-cli

    Self-mode push carries teamai.yaml changes onto the updated default branch

    When the checkout is behind the default branch, teamai push --all now carries the member’s teamai.yaml changes onto the updated default branch. If the member and a teammate changed the same setting, it safely reports a conflict and does not push.

    Usage details

    Who can use it: teamai-cli members using Self-mode when the working checkout is behind the default branch.

    How to use it: Run teamai push --all in the Self-mode checkout; if a conflict occurs, merge the reported keys and push again.

    Tip: Running teamai pull before pushing reduces unexpected conflicts.

  • SettingLoopForge

    Scheduled tasks retain their selected CodeBuddy model

    Scheduled tasks no longer silently switch to the default model at runtime instead of the selected CodeBuddy model. The owner’s model list is refreshed, and if the selected model is unavailable, the task fails with an explicit error.

    Usage details

    Who can use it: LoopForge scheduled tasks; the selected model must be available in the owner’s CodeBuddy model catalog.

    How to use it: Select a CodeBuddy model when creating or editing a scheduled task; keep the model list and key up to date for the task owner.

    Tip: Run a short manual test with the selected model before putting tasks into production.

  • CapabilityLoopForge

    After handoff, native session context is rebuilt for the current mode

    If a new native session is opened after a handoff, the replayed prompt now describes the current phase’s toolset and working mode, rather than the old mode from when the turn began. This reduces incorrect instructions, such as telling the workspace CLI to call a platform tool.

    Usage details

    Who can use it: Mode handoff and fresh-session fallback flows in LoopForge native conversations.

    How to use it: Use the mode handoff or resume flow in the conversation as usual; the current phase prompt is rebuilt automatically.

    Tip: After a handoff, ask the agent to recheck the actual file and tool state.

Watchlist

Kimi K3.1 Pro

46
Moonshot AI · Kimi

Ranks 11th and remains the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. The stable Kimi Code 2.6.2 release fixed slow opening and getting stuck with empty responses in long sessions.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; ranks 12th with 44 points. In preview on Google Agent Platform as `muse-wonder-1.0`.

Claude Haiku 5.5

36
Anthropic · Claude

Anthropic’s fast, low-cost model at $0.10 / $0.50. According to Meta, the new Muse Ember delivers the same score at roughly one-third the task cost.

Muse Ember

36
Meta · Meta AI

Released on November 9. Meta’s smallest Muse model; $0.10 / $0.50, 1M context, available in Meta Model API as `muse-ember-1.0` and on all Muse Code plans.

Grok 4.7

35
xAI · Grok

The current model in the Grok app and API. Grok 4.8 is expected to roll out to some Grok Bot users; it has no official API ID, price, or Artificial Analysis score yet. Grok Bot 1.0 was released on November 11.

Tencent Hy4

35
Tencent · Hy

Available via API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31. Today, November 15, is the last day of free usage in WorkBuddy and CodeBuddy.

Sources