YZ·RADARyz.mikail.net
Daily edition · 14 November 2026

DeepSeek API gets web search and file input; ChatGPT Wallet launches

No changes in the TOP 10 today. Rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are unchanged from November 13: 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 unchanged as well. Kimi K3.1 Pro is in 11th place with 46 points and remains the strongest open-weight model. The LMArena Text page is dated November 11, so it is three days old; there, Gemini 4 Carbon, Gemini 4 Argon, and Muse Wonder lead. ERNIE 5.5 Flash Preview, released by Baidu on November 11, was added to LMArena but did not make the top 10 and does not yet have an Artificial Analysis score.

The biggest developer news of the day came from DeepSeek. On November 13, the company brought web search, file and link reading, and safety checks from its app to the API. The `deepseek-flash` model uses the `deepseek_web_search` tool in the Messages, Responses, and Chat Completions APIs to search the web and for images and provide cited answers; with thinking enabled, the same tool becomes a multi-step deep search. `deepseek_read_url` opens links in a message and adds the page to the context. PDFs, Word, PowerPoint, and images can be uploaded using `POST /files` and passed to the model as plain text or rich layouts (100 MiB and 1,000 pages per file, up to 20 documents per request). `safety_check: {level: high}` checks inputs and outputs and returns a refusal marker if it detects risk. In addition, standalone OCR ($0,225 per 1,000 pages), ASR (20 languages, including Turkish; $0,116 per audio hour, currently half price), and TTS ($0,45 per 10,000 characters, currently half price; no Turkish voice) endpoints are now available. DeepSeek also released an open-source chat client that demonstrates these tools; the client can import DeepSeek App history. Harness v0.2.2-rc.3 also received an experimental importer that does the same job.

OpenAI made no model or API price changes, but took an important step on payments. ChatGPT Wallet began rolling out on November 13 to ChatGPT Pro subscribers in the US aged 18 and older, within dots. A card is added once. Before each purchase, the products, total amount, and card are shown, and approval is given with Face ID or a fingerprint; the merchant receives a payment token generated by American Express, Mastercard, or Visa instead of the actual card number. A Commerce Partner Playbook was also published for businesses. Plus and Pro users can now gift ChatGPT credits from the web. In Japan, credits purchased after November 13 now expire after 180 days instead of 12 months. The stable release of Codex CLI 0.169.0 adds multiple side-chat tabs, the ability to add a custom model provider from the terminal, and a new `--trust` flag for `codex exec`: project settings and MCP servers are no longer loaded in untrusted folders. Nearly all of Anthropic’s 37 entries are from Claude Code 2.1.323 and 2.1.325. Version 2.1.325 fixed a security vulnerability in the decompression of zstd-compressed responses; helper processes started by the sandbox no longer carry input tokens or API keys. A problem causing context errors on every turn in third-party gateway and Microsoft Foundry sessions was fixed. Mod calls to `$.model.complete` and `$.model.classify` are now included in `/cost` and `--max-budget-usd` calculations, while the `pluginDigests` setting is being deprecated.

Gemini 4 Hydrogen continues to roll out at Google: it can be used in mobile and web apps with Firebase AI Logic, and in Android Studio through the Model Providers setting. Gemini users in the US can connect Google Wallet and ask questions about their spending; access is read-only. Google AI Pro and Ultra subscribers can claim monthly Cloud credits directly from AI Studio. Three-legged OAuth for A2A agents and per-user quota limits are now generally available in Gemini Enterprise; API Keys remote MCP servers and Agent Sandbox on GKE are also GA. Meta’s Muse can call businesses on behalf of paid subscribers in the US, while the Muse from Meta Chrome extension accepts tasks from a side panel. At Moonshot, Kimi Work 3.2.27 added Intel Mac support, background execution for long tasks, and the ability to fork or open Side Chat from a running task. Kimi Code 2.6.0 reached stable status; local copying now uses the system clipboard instead of OSC 52 (set `KIMI_CODE_CLIPBOARD_OSC52=1` for the old behavior). Alibaba’s 25 entries are from the Qoder family: Qoder 0.5.11 suggests the next step in an empty input field, Qoder IDE can retry a failed task with a different model, and Qoder CLI 1.1.89 connects directly to enterprise BYOK models. Z.ai’s 29 entries are from ZCode; Z.ai, BigModel, and API Key connections have been combined on a single sign-in screen, and built-in browser pages can be exported as PNG or PDF. LoopForge and teamai-cli at Tencent received security-focused updates: credentials are stripped from repository URLs, and API keys are hidden in SOP logs. The report that Tencent’s live audio model Hy Realtime is now available in Yuanbao is marked unverified because it comes from secondary sources. Xiaomi has MiMo Claw V2.1.2 and MiMo Studio fixes, while xAI has only plugin version updates for Grok Build. A total of 204 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 model `claude-fable-5-2`, $8 / $40, with thinking always adaptive. On November 12, the Claude Meetings beta (Google Meet first) opened to Team and Enterprise plans.Released: 9 November 2026

  2. 02

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    `gemini-4-carbon` is available in the Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, $4 / $20 from December 1). Antigravity’s new Orchestrator mode assigns planning and difficult decisions to Carbon, and routine edits and tests to Hydrogen; Orchestrator (Carbon) is available 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 through 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. In Agents SDK Python v0.24.0-alpha.7, the default reasoning effort for this model increased from `low` to `medium`.Released: 3 September 2026

  5. 05

    GPT-6.1 Sol

    OpenAI · ChatGPT=
    55pts

    The default model on paid ChatGPT plans; API price 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 capable model available on claude.ai’s free plan as well; 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, at a cost of $3,16 per task (compared with $5,48 for Carbon). The API model `gemini-4-hydrogen` is also available on the free tier, with a paid price of $1,50 / $7,50. It can now be used in mobile and web apps with Firebase AI Logic, and in Android Studio through the Model Providers setting (XHIGH is unavailable in Firebase).Released: 12 November 2026

  8. 08

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

    Released on October 26 for ChatGPT Work, Codex, and the API; focused on image and computer use. The management preview for 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; promotional price of $2 / $10 through November 30. In Antigravity, 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 capable Muse model; no public API. This week, the Muse app added the ability 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) is in 11th place.
  • DeepSeek: Web search, deep search, URL reading, file input, and safety checks have been added to the `deepseek-flash` API; OCR, ASR, and TTS endpoints are also available.
  • OpenAI: ChatGPT Wallet is rolling out to Pro subscribers in the US within dots. Each purchase requires biometric approval, and the merchant receives a token instead of the card number.
  • Anthropic: Claude Code 2.1.325 fixed a zstd security vulnerability and separated sandbox helper processes from credentials.
  • Google: Hydrogen has come to Firebase AI Logic and Android Studio; Gemini users in the US can connect Google Wallet.
  • 204 new developments: Anthropic 37, Z.ai 29, Tencent 28, Moonshot 27, Alibaba 25, OpenAI 16, Xiaomi 15, DeepSeek 10, Google 10, xAI 4, Meta 3.

Worth trying today

01

Enable web search and file input in the DeepSeek API

Set the model to `deepseek-flash` and add `deepseek_web_search` to the `tools` list. Keep thinking off for quick questions that need current information; the model will search and provide cited answers. Turn thinking on for research questions; the model will search in several steps and read links itself. If the user message contains a link, also add the `deepseek_read_url` tool. In multi-turn chats, send back the previous turn’s `server_tool_use` and result blocks unchanged. For a PDF or Office file, first upload it with `POST https://api.deepseek.com/files`: use `purpose=text_recognition` for plain text, or `rich_layout_recognition` if charts and tables matter. Add the returned `file_id` to the message. If your app is public, include `safety_check: {level: high}` in the request and handle refusal responses in the interface. Tool charges are in addition to token charges; check the Models & Pricing page.

DeepSeek · DeepSeek
02

Update Claude Code to 2.1.325 and check three things

Update with `claude update` and confirm it is version 2.1.325 with `claude --version`; the zstd security fix requires no configuration. First, a helper program, proxy, or hook running inside the sandbox can no longer read Claude Code’s token or API key from an environment variable; provide any required key through a separate secret mechanism. Second, if you use mods, check costs with `/cost` and set a limit for headless runs with `--max-budget-usd`; model calls made by mods are now included in the total. Third, if you have your own plugin, run `claude plugin validate`, fix any skipped unsafe paths, and plan to remove the `pluginDigests` setting if you use it.

Anthropic · Claude
03

Use the `codex exec --trust` flag correctly in Codex 0.169

Update Codex CLI to 0.169.0 (`npm i -g @openai/codex@latest`). If you use `codex exec` in CI or scripts, note that in an untrusted folder, project configuration and MCP servers are no longer loaded—even if you specify `--sandbox workspace-write`. If you need them in your own repository, add `--trust` to the command (`codex exec --trust "..."`); do not use this flag in external repositories you have not reviewed. If the terminal workspace is open, you can open multiple side-chat tabs and switch between them using the keyboard or mouse. If you use an enterprise gateway, you can now add a custom provider from the TUI and sign in with a device code in an SSH session.

OpenAI · ChatGPT
04

Claim the Cloud credits from your Google AI Pro/Ultra subscription in AI Studio

If you are the primary subscriber to Google AI Pro or Ultra, open Google AI Studio and create your Google Developer Program profile from the credit notice in the left menu or through the Claim flow. Redeem the credits and link your Cloud billing account. The credits are used for paid Gemini API usage; each credit is valid for 12 months from the date it is issued and rolls over from month to month. A minimum $5 prepayment is required for the Gemini API. Track your balance and usage at aistudio.google.com/billing. These credits are a good way to get started with Gemini 4 Carbon or Argon, which require a paid tier.

Google · Gemini

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 · 37

  • SettingClaude Code on the web

    Remote settings queries restricted in cloud and self-hosted sessions

    Settings, permission rules, hooks, and status queries may now return less information or be rejected over a cloud or self-hosted connection. The settings actually applied within the session are unchanged.

    Usage details

    Who can use it: Cloud and self-hosted runner sessions; enterprise security policies.

    How to use it: Use the session’s own `/status`, settings, and permissions views; don’t assume you can read the entire configuration over a remote connection.

    Tip: Write automations to tolerate incomplete remote responses.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    Context overflow compaction fixed in Gateway and Foundry sessions

    Fixed an issue that caused context length errors on every turn in many third-party gateway and Microsoft Foundry sessions. Sessions now continue compacting when appropriate.

    Usage details

    Who can use it: Claude Code sessions using third-party gateways and Microsoft Foundry.

    How to use it: Update to 2.1.325; expect automatic compaction when the context fills up.

    Tip: For long tasks, still manage context proactively with `/compact` or `/autocompact`.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • AvailabilityClaude Code on the web

    Hosted cloud sessions can start without a sandbox

    Claude Code on the web, cloud scheduled tasks, and Cowork cloud sessions no longer refuse to start in a hosted container because of `sandbox.failIfUnavailable`; these containers do not have a command sandbox.

    Usage details

    Who can use it: Claude Code on the web, cloud scheduled tasks, and Cowork cloud sessions.

    How to use it: Start a hosted session; separately assess the session’s hosted security boundaries for steps that rely on a local sandbox.

    Tip: Choose a self-hosted or local environment for tasks that require a sandbox.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • SettingClaude Code computer use

    More risky key combinations blocked during macOS computer use

    macOS computer use now rejects key combinations that hold Control or Option together with Space, as well as more system shortcuts that require permission.

    Usage details

    Who can use it: Sessions using Claude Code computer use on macOS.

    How to use it: When a shortcut is rejected during a computer use request, ask the user to perform the required system action.

    Tip: Design workflows that require explicit user consent instead of automating system shortcuts.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • APIClaude Code Mods

    Mod model calls included in cost and usage metrics

    `/cost`, `$.session.usage()`, `--max-budget-usd`, and cost/token metrics now also count mods’ `$.model.complete` and `$.model.classify` calls.

    Usage details

    Who can use it: Claude Code mods and sessions that track their budgets or usage.

    How to use it: After a mod call, check the total with `/cost` or `$.session.usage()`; use `--max-budget-usd` for headless runs.

    Tip: When evaluating mod costs, track their own model calls as a separate budget item.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • Plugin / IntegrationClaude Code plugins

    Plugin skips unsafe paths and nested skill loading is restricted

    The plugin loader now skips some unsafe paths and links, and `claude plugin validate` reports them. `.claude/skills` directories inside `--plugin-dir` are not automatically loaded in sessions started from a parent directory.

    Usage details

    Who can use it: Teams developing custom plugins, skills, and marketplaces.

    How to use it: Run `claude plugin validate`; start a session from within the plugin, or explicitly move skills to an appropriate plugin location.

    Tip: Distribute plugin packages using real, restricted paths instead of symbolic links.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • APIClaude Code Remote Control

    Remote Control reports its own shutdown and registration failures with different exit codes

    When `claude remote-control` stops on its own, it now exits with 4; when it fails to register, it exits with 69. The normal exit code of 0 is preserved for single-session runs.

    Usage details

    Who can use it: Teams running Remote Control with a service manager.

    How to use it: Treat exit codes 4 and 69 in the service definition as restart or alert conditions.

    Tip: When updating the supervisor configuration, continue to treat the single-session code 0 as success.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    Session credentials isolated from sandbox helper processes

    Programs launched by the sandbox itself, such as its network bridge, no longer carry Claude Code login tokens or API keys in their environment. Internal ripgrep, git, and helper processes also no longer receive the mTLS client key password.

    Usage details

    Who can use it: Enterprise setups using a sandbox, proxy, credential helper, or hook.

    How to use it: Update to 2.1.325; explicitly provide any required credentials to the helper program using a secure secret mechanism.

    Tip: Stop relying on Claude Code environment variables implicitly in credential helpers.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    Security vulnerability fixed in Zstd response decompression

    Claude Code fixed a security issue that occurred when processing zstd-compressed responses. The fix enables safer decompression of compressed server responses.

    Usage details

    Who can use it: Claude Code 2.1.325 and later; all platforms receiving compressed responses.

    How to use it: Update Claude Code to version 2.1.325 or later; no separate setting is required.

    Tip: Don’t delay updates on machines using enterprise setups or gateways.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • Plugin / IntegrationClaude Code MCP tasks

    Background MCP questions can be answered later with Esc

    Pressing Esc on a background MCP task question no longer sends an empty or incorrect answer to the server; the question remains in `/tasks` and can be answered later.

    Usage details

    Who can use it: Claude Code sessions using background MCP tasks.

    How to use it: Press Esc while the question is displayed; later, open `/tasks`, press Enter to show the question again, and answer it.

    Tip: To avoid accidentally answering an MCP question while writing a prompt, submit or clear your draft first.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • SettingClaude Code

    Maximum overload retry delay raised to 300 seconds

    The allowed maximum for `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS` has increased from 64,000 to 300,000 milliseconds.

    Usage details

    Who can use it: Claude Code installations configured for automatic retries on overload errors.

    How to use it: Set the environment variable to no more than `300000`; for example, `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS=300000`.

    Tip: Before setting a long delay on shared infrastructure, account for task timeouts and the CI budget.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code on the web

    Cloud tool search MCP description limit increased to 16,384 characters

    For MCP tool descriptions loaded using tool search in new cloud sessions, the truncation limit has increased from 4,096 to 16,384 characters.

    Usage details

    Who can use it: Claude Code on the web cloud sessions.

    How to use it: Start a new cloud session and call the MCP tool using tool search; longer descriptions are now preserved.

    Tip: Keep MCP tool descriptions concise and task-focused to improve search quality.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • AgentClaude Code

    File and MCP sizing fixed for subagents using a different model

    Subagents with their own model no longer incorrectly size file reads and MCP results using the main session’s model. A fix also prevents the subagent from being incorrectly switched to its planning model when using the `opusplan` or `haiku` setting.

    Usage details

    Who can use it: Claude Code subagents and sessions that use model switching.

    How to use it: Explicitly assign a model to the subagent, then run a file or MCP operation; no additional settings are required.

    Tip: Set the subagent’s model before the task starts to ensure the correct context budget for long results.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • APIClaude Code CLI

    First claude -p call waits for the account cache

    When no account data is cached, `claude -p` no longer sends the first request immediately with the wrong identity; it waits until the account data is ready.

    Usage details

    Who can use it: `claude -p` sessions running in a clean environment or on a new CI worker.

    How to use it: Run `claude -p` in a clean environment; account for the possibility that the first request may be delayed by a few seconds while credentials are prepared.

    Tip: Set the CI timeout to allow for the initial authentication delay.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • FeatureClaude Code VS Code extension

    Cell rerun tip added for Jupyter notebooks

    After Claude reads or edits a Jupyter cell, a tip for rerunning the cell appears in VS Code and similar integrated terminals.

    Usage details

    Who can use it: Claude Code 2.1.325 and later; VS Code and similar integrated terminal surfaces.

    How to use it: Have Claude read or edit a `.ipynb` file, then rerun the cell by following the relevant spinner tip.

    Tip: To verify Claude’s output, rerun only the affected cells after making changes.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    The path forward is made clear when an organization policy blocks a request

    When an organization policy blocks a request, the error message now explains how to proceed: change the message with `/rewind` or run `/clear`.

    Usage details

    Who can use it: Claude Code sessions managed by organization policies.

    How to use it: Follow the suggestion in the block message: run `/rewind` and narrow the prompt, or use `/clear` for new context.

    Tip: Explicitly remove the risky part of the request instead of trying to bypass the policy block.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • Plugin / IntegrationClaude Code plugin marketplace

    Marketplace ref field accepts a full commit ID on a branch or tag

    Marketplace `ref` can now be pinned to any full commit ID on a branch or tag. There is a compatibility warning because clients older than this release cannot clone such a pin.

    Usage details

    Who can use it: Teams managing private plugin marketplaces; Claude Code 2.1.325 and later.

    How to use it: Set the full commit SHA in the `ref` field of the marketplace definition and upgrade older Claude Code clients.

    Tip: For enterprise deployments, pin marketplace refs to signed or verified commits.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • Plugin / IntegrationClaude Code plugins

    MCP plugin bundles can be installed in read-only plugin directories

    `.mcpb` and `.dxt` bundles are now extracted to the plugin directory instead of `.mcpb-cache` inside the plugin directory. This allows read-only plugin directories to be installed; packages fetched from a URL can be downloaded again on the first attempt.

    Usage details

    Who can use it: Plugin marketplace installations that use `.mcpb` or `.dxt`.

    How to use it: Install the plugin with the latest Claude Code; you don’t need to grant write access to the read-only plugin directory.

    Tip: Keep plugin directories read-only to reduce the risk of changes at runtime.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    Approved plan saved to the default folder if the plan directory is unwritable

    If the `plansDirectory` location is unwritable, the approved plan is no longer lost; it is saved to the default plan folder, and a warning indicates the reason and location.

    Usage details

    Who can use it: Claude Code sessions using `plansDirectory`.

    How to use it: Approve the plan; if the custom folder is unwritable, check the default location shown.

    Tip: In enterprise installations, verify in advance that the Claude Code process has write permission for `plansDirectory`.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • RetiredClaude Code plugin marketplace

    The pluginDigests setting is deprecated

    Remote marketplace plugins declared by managed settings are considered the organization’s plugins; the legacy `pluginDigests` field is now deprecated and ignored.

    Usage details

    Who can use it: Organizations that distribute plugin marketplaces through managed settings.

    How to use it: Remove `pluginDigests` from managed settings; manage plugins with `prependPlugins` or `appendPlugins`.

    Tip: For plugin version integrity, use a version range and a verified marketplace commit instead of a digest.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code

    A note explains protected files during sandbox git operations

    If `git checkout`, `git stash`, or an autostash pull inside the sandbox cannot update protected files such as `.claude`, Claude receives a note explaining why.

    Usage details

    Who can use it: Claude Code sessions using a sandbox.

    How to use it: Read the operation note; for protected files, first provide appropriate permissions or a trusted workspace, then rerun the git operation.

    Tip: Back up `.claude` settings before changing them with stash or checkout, and verify that the workspace is trusted.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code CLI

    Shell commands no longer run in the wrong directory when the working directory is deleted

    If the initial working directory is deleted, shell commands no longer silently run in the launch directory; they return a meaningful error and allow the command to be resubmitted.

    Usage details

    Who can use it: Claude Code CLI and sessions where the working directory may later be deleted.

    How to use it: After the error, switch to a valid working directory and resubmit the command.

    Tip: In scripts that delete temporary folders, don’t remove the working directory while a Claude Code session is running.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • CapabilityClaude Code CLI

    Unwanted control sequences filtered from terminal text output

    Terminal text output from `claude -p` no longer passes through control sequences other than formatting sequences. The behavior of pipe and file output has not changed.

    Usage details

    Who can use it: `claude -p` users who receive text output in the terminal.

    How to use it: Run `claude -p` on the latest version; use pipe or file redirection for machine-processed output.

    Tip: Route output through a pipe to keep CI logs safe and stable.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • APIClaude Code CLI

    plugin eval now rejects an empty output directory

    `claude plugin eval --output-dir ''` no longer silently writes the report to the current working directory; it exits with an error.

    Usage details

    Who can use it: Developers who run plugin evaluations through the CLI or CI.

    How to use it: Provide an explicit, non-empty directory path for `--output-dir`.

    Tip: In CI scripts, create the output directory before running the command and use an absolute path.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • SettingClaude Code

    Unrecognized teammateMode values are no longer silently ignored

    An invalid `teammateMode` setting is now reported as a skipped entry in the `claude -p` startup warning and `claude doctor` output, even though it is not applied.

    Usage details

    Who can use it: Claude Code setups using agent team settings.

    How to use it: Check the `claude doctor` or `claude -p` output and replace the unrecognized `teammateMode` value with a valid one.

    Tip: Run `claude doctor` in CI to validate managed settings files against version-compatible values.

    ✓ Verified by official sourceSource: Claude Code v2.1.325 sürüm notları
  • FeatureClaude Code Terminal.app

    Option key behavior explained during Terminal.app setup

    The first-run screen now explains what the “Use Option as Meta key” setting does, how Option affects typing, and how to undo the setting.

    Usage details

    Who can use it: Users launching Claude Code for the first time with macOS Terminal.app.

    How to use it: Read the explanation on the setup screen and confirm your Option key preference.

    Tip: Test your existing shortcuts before changing Option behavior on Turkish or special-character-heavy keyboards.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • FeatureClaude Code

    The exhausted sticker pack no longer opens a misleading browser page

    `/stickers` now says the pack is exhausted instead of opening a page in the browser when the sticker pack has been fully claimed.

    Usage details

    Who can use it: Claude Code users.

    How to use it: Run `/stickers` in Claude Code; if the pack is exhausted, wait for the status shown.

    Tip: Treat this message as a signal to wait for a new distribution, not as an error.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • SettingClaude Code

    Variable for disabling subagent progress summaries documented

    The `CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES=1` setting disables subagent progress summaries and the requests made to generate them.

    Usage details

    Who can use it: Available since Claude Code 2.1.318, formalized in the 2.1.323 notes; headless and interactive sessions.

    How to use it: Set `export CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES=1` before launching Claude Code.

    Tip: Disable it if quota or reducing noise is a priority; leave it enabled if you need visible progress on long tasks.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • SettingClaude Code

    Startup warning added for variable wildcards in Bash permission rules

    A startup warning is displayed for Bash permission rules that use `*` in a variable value; `Bash(PORT=* npm run dev)` is given as an example.

    Usage details

    Who can use it: Claude Code 2.1.323 and later; all installations using Bash permission rules.

    How to use it: Restart Claude Code and check the startup warning; if needed, replace the rule with a narrower pattern.

    Tip: Explicitly restrict commands and parameters instead of allowing variables with wildcards.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • APIClaude Code CLI

    Plugin install JSON output now includes the number of unset configuration options

    The `claude plugin install --json` output now reports the number of `userConfig` options that have not yet been set in the `unsetOptionCount` field.

    Usage details

    Who can use it: CLI and CI workflows that automate plugin installation.

    How to use it: Run `claude plugin install --json <plugin>` and read the `unsetOptionCount` field.

    Tip: If this value is not zero after installation, ask the user to complete the missing plugin settings.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • APIClaude Code CLI

    Customize the exit resume hint with a custom command

    Programs that launch Claude Code can now use `CLAUDE_CODE_RESUME_HINT_COMMAND` to show their own resume command in the exit hint.

    Usage details

    Who can use it: Developers who launch Claude Code with another tool or wrapper program.

    How to use it: Set the `CLAUDE_CODE_RESUME_HINT_COMMAND` variable in the launcher environment to your own resume command.

    Tip: Generate the full command users can copy and run, including any required project or session ID.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • FeatureClaude Code

    Close hint and new shortcut guide added to diff panel

    The diff panel now shows the `/diff to close` hint. The keys for opening and closing the panel are also listed in the `?` shortcut guide.

    Usage details

    Who can use it: Claude Code 2.1.323 and later; terminal users.

    How to use it: Run `/diff` while a file is changed; open the `?` guide to see the shortcuts.

    Tip: Keep the diff panel open and review changes quickly before accepting large changes.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • APIClaude Code Mods

    Mod panels now include an isPromptCovered indicator

    The mod panel’s drawing props now include the `isPromptCovered` field. Mods can detect when a dialog or the area above it covers the prompt.

    Usage details

    Who can use it: Claude Code mod developers.

    How to use it: Read `isPromptCovered` from the mod panel’s drawing props to adjust the position or visibility of panel content.

    Tip: Improve accessibility by keeping keyboard focus out of the panel when the prompt is covered.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • SettingClaude Code

    Startup notice names the Perforce setting when it is not enabled

    When Perforce settings are detected but the mode is off, Claude Code displays a one-time warning naming the `CLAUDE_CODE_PERFORCE_MODE` variable.

    Usage details

    Who can use it: Claude Code installations that use Perforce.

    How to use it: Start Claude Code in an environment with Perforce settings; configure the `CLAUDE_CODE_PERFORCE_MODE` variable if needed.

    Tip: In team environments, enable Perforce mode consistently with managed settings or leave it explicitly disabled.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları
  • Plugin / IntegrationClaude Code

    An informative warning appears when a plugin language server fails to start

    If a plugin’s language server cannot start, Claude Code now displays a warning identifying the server and what to check.

    Usage details

    Who can use it: Installations that use Claude Code plugins.

    How to use it: After installing a plugin, check the startup output or the result of `claude plugin list`; address the server and configuration mentioned in the warning.

    Tip: Pin language server dependencies to the versions specified in the plugin documentation.

    ✓ Verified by official sourceSource: Claude Code v2.1.323 sürüm notları

Google · Gemini

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

New in this edition · 10

  • Plugin / IntegrationGemini Enterprise

    Three-legged OAuth for Gemini Enterprise A2A agents is now GA

    Gemini Enterprise has made 3-legged OAuth support generally available for A2A agents with the Agent Identity auth manager. Agents built with ADK and running in Agent Runtime can access MCP servers, Google Cloud services, and third-party APIs on behalf of a user, with the signed-in user’s consent.

    Usage details

    Who can use it: Gemini Enterprise A2A agents; deployments using ADK and Agent Runtime.

    How to use it: Select the 3-legged OAuth configuration when registering the A2A agent; provide the OAuth client ID/secret, authorization and token URIs, and required scopes to the Agent Identity auth manager. Gemini Enterprise starts the flow when user consent is needed, validates the callback at https://vertexaisearch.cloud.google.com/validateUserId, and resumes the conversation.

    Tip: Request only the OAuth scopes required for access on behalf of the user, and keep the end-user OAuth identity separate from the service-to-service IAM identity.

  • Pricing / PlanGemini Enterprise

    Single-user quota limit for AI developer tools is now GA in Gemini Enterprise

    A setting that limits how much a single user can consume from the shared quota pool is now GA in Gemini Enterprise Standard and Plus; the limit also applies to AI developer tools, including Google Antigravity. For these tools, the multiplier applies to the user’s seven-day quota cycle and resets with each user’s cycle.

    Usage details

    Who can use it: Gemini Enterprise Standard and Gemini Enterprise Plus editions; projects with administrator permissions.

    How to use it: In the Google Cloud console, open Gemini Enterprise > Usage & spend > Usage; select the location and edition. Under Feature usage > Limit any single-user’s quota, turn on Enabled, enter a Multiplier value, and click Save changes.

    Tip: Start with a low multiplier, then monitor how quickly the shared pool is depleted on the Usage & spend screen.

  • Plugin / IntegrationGemini

    Connect Google Wallet to Gemini for personal finance insights

    All Gemini app users in the US can connect Google Wallet to get personal finance answers based on their spending habits. Users can also connect bank and credit card accounts through Plaid if they choose; Gemini can only read data and cannot move, manage, or deposit money.

    Usage details

    Who can use it: Web, Android, and iOS versions of the Gemini app in the US; accounts with access to Personal Intelligence.

    How to use it: In Gemini, open Profile > Personal Intelligence > Connected Apps and enable the Google Wallet connection; then use the @Google Wallet tag in a chat to ask about spending, budgets, or subscriptions.

    Tip: For your first query, specify a period and a goal; for example, ask it to review the last quarter and create a three-month savings budget.

  • Pricing / PlanGoogle AI Studio

    Google AI Pro and Ultra Cloud Credits can be claimed from AI Studio

    Google AI Pro and Ultra subscribers can create a Google Developer Program profile, link billing, claim monthly Google Cloud Credits, and track their balance without leaving Google AI Studio. Credits can be used for paid Gemini API usage; each credit is valid for 12 months from issuance and rolls over monthly, and a minimum $5 prepayment is required for the Gemini API.

    Usage details

    Who can use it: Primary subscriber to Google AI Pro or Google AI Ultra; active Google Developer Program profile and Cloud billing setup.

    How to use it: In Google AI Studio, open the credit notice in the left menu or the Claim flow; create a Developer Program profile, Redeem the credit, and link Cloud billing. Track your balance and usage at aistudio.google.com/billing.

    Tip: The system uses the credits that expire soonest first; still, regularly track spending and balances for paid API keys.

  • AvailabilityAgent Sandbox on GKE

    Agent Sandbox on GKE is now GA for isolated AI agent Pods

    Agent Sandbox on GKE is now GA for isolated, state-preserving single-replica workloads and runs on all supported GKE versions. It is designed to run untrusted code generated by AI agents in separate Pods, use prewarmed environments, and isolate workloads that require GPUs/TPUs or a full PodSpec.

    Usage details

    Who can use it: Supported GKE versions; Autopilot or Standard clusters. There is no additional charge for the software, but GKE charges apply for the resources it creates.

    How to use it: Prepare a GKE cluster and gVisor or microVM isolation; install the Agent Sandbox controller manifest or Helm chart. Then request agent environments using the SandboxTemplate, SandboxWarmPool, and SandboxClaim resources.

    Tip: When running untrusted code, do not start a Pod without specifying gVisor or microVM; otherwise, there is no kernel isolation.

  • Plugin / IntegrationAPI Keys API

    API Keys remote MCP server is now GA for AI applications

    Google Cloud’s API Keys remote MCP server is now GA. Antigravity, ChatGPT, Claude, and custom AI applications can use this server to create, inspect, restrict, and manage the lifecycle of API keys in Google Cloud projects.

    Usage details

    Who can use it: Google Cloud projects with the API Keys API enabled; AI applications or agents with OAuth 2.0 and the required IAM roles.

    How to use it: Add a remote MCP server in your AI client: URL https://apikeys.googleapis.com/mcp, transport Streamable HTTP, OAuth scope https://www.googleapis.com/auth/apikeys. Grant the MCP Tool User and API Keys Admin or Viewer roles.

    Tip: Use a separate Google Cloud identity for agents; do not use API keys for MCP authentication.

  • AvailabilityFirebase AI Logic

    Firebase AI Logic brings Gemini 4 Hydrogen to mobile and web apps

    Firebase AI Logic has made Gemini 4 Hydrogen directly available to mobile and web apps through the Gemini Developer API free tier or the Agent Platform Gemini API. On Agent Platform, the model is available for a limited time; HIGH is the default thinking level in Firebase, and XHIGH, thinkingBudget, and long decode are not supported.

    Usage details

    Who can use it: Mobile and web apps using the Firebase AI Logic SDK; Gemini Developer API free tier or Agent Platform Gemini API.

    How to use it: Add the Firebase AI Logic SDK and use gemini-4-hydrogen as the model ID; in production, manage the model name dynamically with Firebase Remote Config or server prompt templates.

    Tip: Set thinkingLevel to MEDIUM or LOW to reduce latency and usage; do not send thinkingBudget.

  • AvailabilityAndroid Studio

    Gemini 4 Hydrogen is available to developers in Android Studio

    Google announced that coding, debugging, interactive web development, and multimodal workflows are available in Android Studio with Gemini 4 Hydrogen. This is a new developer surface in Android Studio, separate from the known general model GA listing.

    Usage details

    Who can use it: The current stable channel of Android Studio and releases published in the previous 10 months; individual Gemini users, users with API keys, and Gemini Enterprise business users.

    How to use it: In Android Studio, open File > Settings > Tools > AI > Model Providers; add a Gemini API key from Google AI Studio, select gemini-4-hydrogen from the Available models list, and click Apply.

    Tip: If you use your own API key, monitor model usage on the Usage and Limits screen in AI Studio; do not put the key in source code.

  • CapabilityModel Armor

    Model Armor Latest improved detection of unintended commands and jailbreaks with v4

    The v4 filter indicated by the Latest alias in Model Armor delivers higher recall for indirect prompt injection, tool calls, conversational inputs, and jailbreak attacks; the false-positive rate for conversational inputs is also lower. The update strengthens prompt and response security in Gemini Enterprise Agent Platform integrations.

    Usage details

    Who can use it: Model Armor templates and Google Cloud AI-agent workloads; regions where v4 is supported and Gemini Enterprise Agent Platform integrations.

    How to use it: Select Latest as the filter version selector in the Model Armor template to automatically get current protections; if you need immutable behavior, select v4 as a fixed version and provide the template ID as model_armor_config in the Gemini call.

    Tip: First evaluate Latest against real traffic samples in pre-production; use a fixed version if you need immutable behavior in production.

  • AvailabilitySecure Web Proxy

    Secure Web Proxy opens an Agent Gateway preview for AI agents

    Secure Web Proxy can now be used as a self-managed Agent Gateway for traffic from AI agents running on GKE and Cloud Run to LLMs, MCP tools, remote agents, and external APIs. The Preview feature provides centralized access policies, identity controls, Model Armor protections, and observability within a VPC.

    Usage details

    Who can use it: Preview; Google Cloud projects running AI agents on GKE or Cloud Run whose access requests are approved.

    How to use it: Fill out the Preview access form; associate the Secure Web Proxy instance with Agent Gateway, configure IAP or Model Armor extension bindings and MCP authorization policies, then route the agents’ external traffic through the proxy.

    Tip: Use authorization policy to restrict MCP tool calls, and Model Armor binding to sanitize prompts and responses.

OpenAI · ChatGPT

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

New in this edition · 16

  • FeatureChatGPT Wallet

    ChatGPT Wallet lets dots make purchases with user approval

    ChatGPT Wallet lets dots make purchases on a user’s behalf; for each transaction, the items, total, and card are reviewed and approved with a biometric security check. A payment token created for the transaction is used instead of the actual card details.

    Usage details

    Who can use it: Initial rollout in the US for ChatGPT Pro users aged 18 and older and people with eligible US-issued cards; the feature is available in dot conversations and is rolling out gradually. Eligible American Express, Chase, Mastercard, and Visa cards are supported.

    How to use it: Create a dot; on mobile, go to Settings > Wallet > Add Card, or ask the dot to add a card and complete the secure form. During checkout, review the details and approve with Approve and Face ID/Touch ID.

    Tip: Do not enter card numbers, security codes, or verification codes in chat; use only the secure card form and check recurring-payment terms before approving.

  • Pricing / PlanOpenAI Service Credits

    Credits purchased in Japan after November 13 expire after 180 days

    According to OpenAI’s current credit terms, Service Credits purchased or granted in Japan on or after 2026-11-13 expire after 180 days. The general 12-month validity rule does not apply to the Japan exception.

    Usage details

    Who can use it: ChatGPT usage credits, Business credits, gift or redemption credits, and prepaid API credits in Japan; eligibility and product terms for the relevant account or workspace apply.

    How to use it: Check your balance and purchase date in ChatGPT under Settings > Usage, in Codex under Usage & Billing, or in API Platform under Billing overview; use new credits in Japan within 180 days.

    Tip: Set automatic reloading only for an amount you can use within 180 days.

  • FeatureChatGPT

    ChatGPT credits can be gifted by Plus and Pro users

    Eligible personal Plus or Pro accounts and the Personal workspace of a Business account can purchase ChatGPT credits and gift them using a single-use claim link. Recipients redeem the link on an eligible personal account within 30 days; gifted credits are valid for 12 months after redemption.

    Usage details

    Who can use it: Phased rollout; purchases are available on ChatGPT web for personal Plus/Pro accounts or in the Personal workspace of a Business account. Recipients can have personal Free, Go, Plus, or Pro accounts, or use the Personal workspace of a Business account; Enterprise and Business workspaces are not eligible.

    How to use it: On the web, switch to an eligible account or Personal workspace, open the credit-gifting flow, select an amount, complete payment, then send the single-use claim link by email or share it through a secure channel. The recipient opens the link, selects the correct personal account, and chooses Redeem.

    Tip: Protect the claim link like a password; anyone else with the link could redeem the credits first.

    ✓ Verified by official sourceSource: Gifting credits in ChatGPT
  • CapabilityCodex Cloud

    You can now add images to Codex Cloud conversations from the terminal

    Images can be sent to Cloud conversations in macOS and Linux terminal installations using an optional Cloud connection. This is a new image-attachment capability, separate from the model selection and conversation archiving features in the same release.

    Usage details

    Who can use it: Codex CLI 0.169.0; users on macOS and Linux with an opt-in Cloud connection.

    How to use it: Update Codex CLI to 0.169.0, enable the Cloud connection, and attach or send an image in a Cloud conversation from the terminal.

    Tip: Reduce the need for explanations by adding error screenshots, design references, or chart outputs directly to the task context.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • SettingCodex CLI

    Codex exec keeps project settings and MCP disabled in untrusted directories

    codex exec does not enable project configuration or MCP servers in untrusted directories, even when --sandbox workspace-write is used. You can use --trust when invoking codex exec or codex review for a project you explicitly trust.

    Usage details

    Who can use it: Codex CLI 0.169.0; codex exec and codex review commands.

    How to use it: Add --trust to the command for a trusted repository: codex exec --trust ... . Do not use this flag in directories you do not trust.

    Tip: Verify the repository’s source before running code or MCP configuration; do not make --trust a general-purpose default.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • CapabilityCodex CLI

    Codex TUI adds multiple side chat tabs and reopening

    The optional terminal workspace now supports multiple side chat tabs. You can switch between tabs using the keyboard or mouse, and saved local side chats can be reopened after restarting Codex.

    Usage details

    Who can use it: Codex CLI 0.169.0; installations using the optional terminal workspace.

    How to use it: Update Codex CLI to 0.169.0 and enable the opt-in terminal workspace; select side chat tabs using the keyboard or mouse.

    Tip: Keep research or reviews separate from the main task in side chats to reduce context mixing.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • Plugin / IntegrationCodex CLI

    You can now add and edit custom model providers from the Codex TUI

    Custom model providers can be added and edited from the Codex terminal. When custom gateway onboarding is enabled, supported gateways can be accessed over SSH or by using a device code without a local browser.

    Usage details

    Who can use it: Codex CLI 0.169.0; installations with custom gateway onboarding enabled that use a supported gateway.

    How to use it: Update Codex CLI to 0.169.0, enable custom gateway onboarding, and add or edit a provider from the terminal during the sign-in flow; use the device code option in SSH environments.

    Tip: On remote machines without browser access, signing in with a device code is safer and more practical than copying local credentials.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • APIOpenAI Node SDK

    Node SDK updates Agents, Live, Images, and admin types

    openai-node 7.57.0 updates the Agents, Live, Images, and admin API types. This release makes new fields in the API surface available in the TypeScript/JavaScript SDK.

    Usage details

    Who can use it: API developers using OpenAI Node SDK 7.57.0.

    How to use it: Update the package: npm install openai@7.57.0 . If you use TypeScript, rerun type checks for agent, Live, image, and admin calls.

    Tip: In applications using exhaustive unions or admin APIs, check for build errors and serialization of new fields immediately after upgrading.

    ✓ Verified by official sourceSource: OpenAI Node SDK 7.57.0 release notes
  • CapabilityCodex CLI

    Codex cleanly closes interrupted turns

    Interrupted turns now close cleanly; delayed cancellations or file-approval cancellations no longer accidentally interrupt the next turn.

    Usage details

    Who can use it: Codex CLI 0.169.0; long-running terminal turns that require approval.

    How to use it: Update Codex CLI to 0.169.0; after canceling a turn, start the next one as usual.

    Tip: If a pending approval prompt remains visible after cancellation, check the terminal state before starting a new turn.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • CapabilityCodex CLI

    Codex plan handoff carries the full approved plan into a new session

    The “Yes, clear context and implement” flow now transfers the entire approved plan to the new session. If the transfer fails, the current session remains open.

    Usage details

    Who can use it: Codex CLI 0.169.0; sessions using the clear-context-and-implement flow.

    How to use it: After approving a plan, choose “Yes, clear context and implement” in Codex; check the plan transfer in the new session against the first prompt.

    Tip: After the transfer, review the new session’s plan summary and continue execution if the file or scope changes are correct.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • Plugin / IntegrationCodex CLI

    Codex TUI no longer waits on an MCP server that was not reported

    The terminal no longer stays busy waiting for MCP servers that the shared app server did not report.

    Usage details

    Who can use it: Codex CLI 0.169.0; terminal sessions using an app server and MCP.

    How to use it: Update Codex CLI to 0.169.0; if the issue persists, refresh the app server and MCP server list, then check the configuration.

    Tip: Verify that the shared app server and local MCP configuration use the same server list.

    ✓ Verified by official sourceSource: Codex 0.169.0 release notes
  • APIOpenAI Agents Realtime JS SDK

    Agents Realtime JS SDK fixes metadata and manual interruption behavior

    @openai/agents-realtime 0.22.0 preserves limited JSON usage metadata and OpenAI wire types; narrowing may be required for Realtime image content, MCP descriptions, schema/tracing metadata, and transcription breakdown types. Saved states use schema 1.23 for compatibility and schema 1.24 for rich metadata or new nested runs; manual interruption also cancels the active WebSocket response before audio starts.

    Usage details

    Who can use it: JavaScript/TypeScript developers using @openai/agents-realtime 0.22.0.

    How to use it: Update the package: npm install @openai/agents-realtime@0.22.0 . Run the TypeScript build and check Realtime content, MCP, tracing, transcription, and saved-state flows.

    Tip: Before migrating saved sessions in production, test the migration by restoring both older compatible states and new states with rich metadata.

  • SettingChatGPT Enterprise ve Edu

    Global admins can change eligible users’ email addresses in Admin Console

    Global admins can change the email addresses of eligible Enterprise and Edu users who are not managed by SCIM in Admin Console. This makes address updates easier for manually managed users outside SCIM.

    Usage details

    Who can use it: ChatGPT Enterprise and Edu; global admins; eligible users outside SCIM.

    How to use it: Go to user management in Admin Console and update the eligible member’s email address; for accounts managed by SCIM, make the change through the identity provider.

    Tip: Check in advance that the new address is not linked to another OpenAI account and that the relevant domains are verified.

  • APIOpenAI Python SDK

    Python SDK expands support for agent, image, Live, and external keys

    openai-python 3.51.0 includes type and surface updates that expand API support for agent, image, Live, and external key features. The release notes describe these four areas collectively.

    Usage details

    Who can use it: API developers using OpenAI Python SDK 3.51.0.

    How to use it: Update the Python package: pip install -U openai==3.51.0 . Then revalidate the agent, image, Live, or external key API types you use against the new version.

    Tip: Before deploying to production, run type checks and a brief regression test for the relevant agent, image, Live, and external key integrations.

    ✓ Verified by official sourceSource: OpenAI Python SDK 3.51.0 release notes
  • APIOpenAI Ruby SDK

    Ruby SDK updates Live, Agents, images, and external key APIs

    openai-ruby 0.124.0 updates the Live, Agents, images, and external keys areas. The release notes publish these changes collectively as an API surface update.

    Usage details

    Who can use it: API developers using OpenAI Ruby SDK 0.124.0.

    How to use it: Update the Ruby gem: gem install openai -v 0.124.0 . Build and test your application’s Live, Agents, image, and external key calls with the new version.

    Tip: Run type and request-body regression tests for existing production calls before recording the gem update in your lockfile.

    ✓ Verified by official sourceSource: OpenAI Ruby SDK 0.124.0 release notes

Meta · Meta AI

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

New in this edition · 3

  • AvailabilityWhatsApp Manager

    Alipay verification for Chinese businesses enters limited testing in WhatsApp Manager

    WhatsApp Manager has launched a limited test allowing some businesses registered in mainland China to verify their business with Alipay. Partners do not need to submit integration or access requests; document verification remains available.

    Usage details

    Who can use it: Businesses with a portfolio registered in mainland China that are not yet verified and have been selected for Meta’s random test group. The option appears in the Overview and Messaging limits surfaces in WhatsApp Manager; the Messaging limits card shows it when the messaging limit is 250.

    How to use it: Open WhatsApp Manager > Overview > Business Status > Verify your business with Alipay or Account tools > Messaging limits > Verify with Alipay. Enter the 18-character Unified Social Credit Code, the legal representative’s name, and their Chinese ID number exactly as registered; then select Scan to verify and scan the QR code using the representative’s own Alipay app.

    Tip: If the Alipay option isn’t visible, check whether the business is registered in China, is unverified, and is in the test group; you can’t request to join the test group, and there is no API for this verification.

  • CapabilityMuse

    Muse calls businesses on behalf of paid subscribers in the US

    Muse now lets paid subscribers in the US call businesses with publicly listed phone numbers on their behalf. Meta’s chief AI officer announced that the feature can be used for tasks such as calling cable companies, the DMV, insurers, or restaurants.

    Usage details

    Who can use it: Paid Muse subscribers in the US; Muse subscriptions are still in limited testing and are not available in every region.

    How to use it: In the Muse app or web version, clearly specify in the chat the business’s publicly listed number and what you need done; check the result and any actions taken when the call is complete.

    Tip: Include the business name, phone number, goal, and acceptable alternatives in the call request; separately verify critical outcomes such as reservations or charges.

  • Plugin / IntegrationMuse from Meta

    Muse from Meta Chrome extension released; tasks can be assigned from the side panel

    Meta has released an extension that brings Muse to the Chrome side panel on the Chrome Web Store; the store listing shows that version 2.0.65 was updated on November 12. Users can continue existing conversations, delegate tasks on web pages, use connected apps, and approve critical actions.

    Usage details

    Who can use it: Available through the Chrome Web Store on Chrome desktop; provided by Meta and used with a Muse account. Official listings do not specify any separate country or subscription restrictions.

    How to use it: Open the ‘Muse from Meta’ page on the Chrome Web Store and select Add to Desktop; open the extension from the Chrome toolbar, sign in with a Muse account, and start a chat or task from the side panel.

    Tip: Decide in advance which accounts you want Muse to act on; monitor its actions in the open browser window and don’t proceed without approving critical steps such as purchases or sending messages.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 27

  • FeatureKimi Work

    Fork and Side Chat are available while a task is running

    You can now fork a session and start Side Chat while a task is running.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; sessions with a running task.

    How to use it: While the task is running, open the session actions and select Fork or Side Chat to work in a separate context without stopping the main task.

    Tip: Try alternative solutions or follow-up questions in Side Chat without disrupting the main task’s context.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • APIKimi Code Server API

    Undo API accepts turn_id and returns removed_turn_ids

    `POST /sessions/{id}:undo` now accepts `turn_id` and an optional `agent_id` instead of a page or message count; the response returns a `removed_turn_ids` list of deleted turns.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; clients managing session history through the Server API.

    How to use it: Use the `{"turn_id":"...","agent_id":"..."}` body in the `POST /sessions/{id}:undo` request and update local history using the `removed_turn_ids` values in the response.

    Tip: Don’t rely on the old `count` or message-page response in your client; remove deleted turn IDs from local state deterministically.

  • CapabilityKimi Work

    Long-running tasks can run in the background without blocking the main flow

    Kimi Work can now run long-running tasks in the background without blocking the main process.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; users running long-running tasks with the desktop Agent.

    How to use it: Start a long-running task in Kimi Work and use the background execution option in the task flow; the main conversation remains available.

    Tip: Move tasks such as file processing and multistep research to the background, then continue asking questions in the main session.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • CapabilityKimi Code CLI

    Skill content over 50,000 characters is moved to a file

    When the model or a nested call loads a Skill over 50,000 characters, it provides a context-limited preview, a truncation note, and an `output_path`. The full content is preserved so it can be read in pages using Read or searched using Grep.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; users of long Skills through TUI, print, Desktop/Web, VS Code, ACP, or SDK.

    How to use it: Call the long Skill as usual; read the `output_path` value in pages with Read or search the full content with Grep.

    Tip: Instead of loading very long Skills directly into the model, start with the summary preview and read only the necessary sections through `output_path`.

  • FeatureKimi Work

    Dragging files onto the desktop pet opens Launcher and attaches them

    When files are dropped onto the desktop pet, Launcher opens automatically and the files are attached.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; desktop app users who use the desktop pet.

    How to use it: Drag and drop one or more files onto the Kimi Work desktop pet; check the attachments in the Launcher that opens, then send the message.

    Tip: Dragging together the files needed for the same task reduces the steps for adding them individually in Launcher.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • FeatureKimi Code CLI

    Activity bar shows current and previous turn durations

    While the model is generating a response, the activity bar shows how long the current turn has taken. When the turn is complete, the duration and completion time of the last turn are also displayed.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; TUI activity bar.

    How to use it: Update the CLI to version 2.6.0 or later and run an Agent turn; duration information appears automatically in the activity bar.

    Tip: For long tasks, track turn duration to decide earlier whether to split the prompt or narrow the task scope.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • APIKimi Code SDK

    Kimi Code SDK 0.23.1 released

    The `@moonshot-ai/kimi-code-sdk` package version has been upgraded to 0.23.1 to synchronize with the Kimi Code 2.6.0 release.

    Usage details

    Who can use it: Developers using the SDK through npm; `@moonshot-ai/kimi-code-sdk` 0.23.1.

    How to use it: Run `npm install @moonshot-ai/kimi-code-sdk@0.23.1` in your package manager and apply the changes to the Server API contract in your client.

    Tip: Before updating in production, test the `undo`, prompt status, and WebSocket `agent.status` handlers against the 2.6.0 contract.

  • AvailabilityKimi Work

    Kimi Work now supports Intel (x86) Macs

    Kimi Work 3.2.27 adds macOS devices with Intel/x86 processors to the supported platforms.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; Intel/x86 macOS devices.

    How to use it: Update Kimi Work to version 3.2.27 or later and restart it on your Intel Mac.

    Tip: If no update appears on your Intel Mac, check both the in-app updater and the official download page.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • FeatureKimi Work

    Notifications and quick actions added to the macOS menu bar

    The macOS menu bar integration provides unread conversation alerts and the New Chat and Open Kimi quick actions.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; macOS users.

    How to use it: Leave Kimi Work running on macOS, then open the Kimi icon in the menu bar to view unread alerts or select the New Chat/Open Kimi actions.

    Tip: Starting a new conversation from the menu bar instead of keeping the app open in its main window speeds up frequently used quick tasks.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • Plugin / IntegrationKimi Work

    One-click environment setup improved for mobile Plugin tasks

    The Plugin workflow in the mobile app helps complete environment setup with one click when it detects a related task.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; mobile app Plugins using related tasks.

    How to use it: Open the related task in the mobile app; when the task is detected, confirm the one-click environment setup prompt.

    Tip: Completing environment setup the first time reduces the need to prepare the same Plugin again for later tasks.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • FeatureKimi Code CLI

    Activated cards hidden for Skills loaded by the model

    In fullscreen view, Skills loaded by the model itself are no longer shown as unnecessary Activated skill cards. Indicators for Skills activated by the user are preserved.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; fullscreen TUI.

    How to use it: Select `/settings` → TUI mode → Fullscreen and restart Kimi Code; Skills loaded by the model will work without showing cards in the summary stream.

    Tip: To reduce noise in long transcripts, use fullscreen view together with an appropriate folding density.

  • Plugin / IntegrationKimi Work

    Plugins can be shared between Work and Chat modes

    Plugins installed or available in Kimi Work can now be used across Work and Chat modes.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; accounts using Plugins in Chat and Work modes.

    How to use it: Install and enable the Plugin from Plugin Center, then use it from the Plugin list in Work and Chat modes on the same account.

    Tip: Confirm that a single shared installation is enabled instead of installing the same Plugin separately in both modes.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • CapabilityKimi Work

    Projects can no longer be saved in the Kimi installation folder

    Kimi Work prevents projects from being saved in the Kimi installation directory to avoid file loss during update cleanup.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; desktop Project users.

    How to use it: When creating or saving a Project, choose a separate user workspace folder instead of the Kimi installation directory.

    Tip: Keeping projects separate from the app folder makes them easier to back up and move independently of updates.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • CapabilityKimi Work

    Remote Control pairing and file previews are more reliable

    The Remote Control pairing flow and remote file preview reliability have been improved.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; devices using Remote Control and file previews.

    How to use it: Update Kimi Work, pair Remote Control again, and use previews from remotely opened file cards.

    Tip: After pairing, check that the remote preview opens the correct file before editing large or critical files.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • APIKimi Code Server API

    Prompts that fail to start a turn are now marked failed, not queued

    If prompt submission fails before starting a turn, the API now reports the status as `failed`; the previous `queued` status has been removed.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; Server API clients and integrations that track prompt status.

    How to use it: Track prompt status in the API response or WebSocket conversation stream; when `failed` is returned, handle the error and retry instead of treating it as queued.

    Tip: Don’t wait only for a `queued` → `running` transition in your client; explicitly handle `failed` as a terminal error and decide whether to retry.

  • APIKimi Code Server API

    WebSocket agent status changes to compacting during compaction

    The WebSocket `agent` frame reports `status: compacting` while context compaction is underway; this period is no longer reported only as `running`.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; WebSocket clients that listen for `agent` frames.

    How to use it: Handle `agent.status` values `idle`, `running`, and `compacting` separately in your WebSocket client; show users an appropriate waiting state during `compacting`.

    Tip: Do not treat `compacting` as an error or idle state; use it to queue new prompts or show users that they should wait.

  • SettingKimi Code CLI

    Local copying uses the system clipboard by default

    In environments other than SSH and mosh, copying now writes only to the system clipboard by default; sending OSC 52 through the terminal has been removed from the default flow. The old behavior can be restored with `KIMI_CODE_CLIPBOARD_OSC52=1`.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; local, SSH, or mosh terminal environments.

    How to use it: To force the old OSC 52 behavior, set `export KIMI_CODE_CLIPBOARD_OSC52=1` before running; no additional configuration is needed for normal local use.

    Tip: If you have clipboard issues in a local terminal, fix the system clipboard utility first; enable OSC 52 only in environments where the terminal requires it.

  • CapabilityKimi Code CLI

    Scan limits and streaming previews added to `@` file completion

    `fs:suggest` now performs a bounded scan, returns up to 20 results, and offers optional streaming previews. When match quality is equal, regular files are ranked first.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; users completing file paths with `@`.

    How to use it: Update the CLI to 2.6.0 and use file path completion by typing `@` in Composer/TUI; results are limited automatically.

    Tip: In large repositories, entering a full folder path narrows the search and helps you find the right file faster within the 20-result limit.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Code CLI

    Concurrent use of the same session in two app instances prevented

    Multiple app instances can no longer restore the same session and run pending task turns in parallel.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; users opening the same session in multiple CLI, Desktop, or Web instances.

    How to use it: Update CLI or Desktop to 2.6.0; when opening the same session in a second instance, use the product’s session ownership and restoration behavior.

    Tip: Using fork or Side Chat instead of opening the same session in two clients reduces concurrent task conflicts.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Code CLI

    Elapsed time shown in the live thinking header

    In the fullscreen TUI, live thinking is now shown with a highlighted Thinking header and elapsed time instead of a spinner.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; TUI sessions with thinking enabled.

    How to use it: Start a CLI turn with a model that supports Thinking; watch the Thinking header in the fullscreen transcript while the model thinks.

    Tip: Use the thinking duration to compare against response latency and adjust `reasoning_effort` or task scope.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Code CLI

    TUI stays responsive with many expanded tool outputs

    A performance issue that caused the terminal interface to stop responding when many tool outputs were opened with `Ctrl-O` has been fixed.

    Usage details

    Who can use it: Kimi Code CLI 2.6.0 and later; users inspecting expanded tool output in the TUI.

    How to use it: Update the CLI to 2.6.0 and open tool output with `Ctrl-O`; the new rendering behavior is used without requiring additional settings.

    Tip: When reviewing large outputs, opening only the necessary tool cards also helps keep the terminal readable.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Code CLI

    Workspace operations are faster with many saved Projects

    The read load causing workspace operations to slow down when many Projects were saved has been reduced.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; users with many Project or workspace records.

    How to use it: Upgrade the CLI to 2.6.0; workspace operations use the new behavior without requiring additional configuration.

    Tip: Regularly cleaning up unused workspace records in large projects helps you get more out of the new speed improvements.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Work

    Global search recall and relevance have been improved

    Global search has been improved to retrieve more relevant results than the previous search feature and match results to queries more effectively.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; users searching project, Plugin, Skill, and conversation content.

    How to use it: Open the global search field in the Kimi Work sidebar and enter a query for project, Plugin, Skill, or conversation content.

    Tip: Start your search with a more distinctive phrase that includes a project or Plugin name instead of a single generic word.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27
  • APIKimi Code Server API

    Session profile API ignores reserved metadata keys

    The Session profile endpoint no longer applies the reserved keys `cwd`, `goal`, `parent_session_id`, and `child_session_kind`; these fields are reserved for the session creation flow.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; API clients using the session profile endpoint.

    How to use it: Do not send reserved keys in profile update requests; provide the working directory, goal, and parent/child information in the session creation configuration.

    Tip: If you use these keys in your custom metadata schema, rename them; otherwise, their values will be silently ignored.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • APIKimi Code Server API

    WebSocket heartbeat disconnects during server stalls and clock jumps fixed

    A condition that caused WebSocket connections to close due to a false heartbeat timeout when the server briefly stalled or the system clock jumped has been fixed.

    Usage details

    Who can use it: Kimi Code 2.6.0 and later; Server API WebSocket connections.

    How to use it: Upgrade the server and client to versions compatible with 2.6.0; use your existing reconnect flow if a connection drops.

    Tip: Before treating a client-side heartbeat disconnect as a permanent server error, check reconnection and session status.

    ✓ Verified by official sourceSource: Kimi Code 2.6.0 GitHub release
  • CapabilityKimi Work

    Paths that differ only by letter case no longer merge on Windows

    An issue that caused paths differing only in letter case to be incorrectly merged into the same path on Windows has been fixed.

    Usage details

    Who can use it: Kimi Work 3.2.27 and later; Windows files and Project workspaces.

    How to use it: Upgrade Kimi Work to version 3.2.27 or later; file paths will be treated separately automatically in the new version.

    Tip: In Windows workspaces, use clear, unique names instead of creating folders with the same name but different letter casing.

    ✓ Verified by official sourceSource: Kimi Work Release Notes — 3.2.27

xAI · Grok

Grok 4.7
Full history (170) →

New in this edition · 4

  • Plugin / IntegrationGrok Build

    Expo Grok Build plugin updated to version 1.14.5

    The pinned version of the Expo plugin in Grok Build’s official plugin marketplace was upgraded from 1.14.3 to 1.14.5. The plugin directory in the catalog and the commit pin were also updated; the official commit does not provide a detailed feature list.

    Usage details

    Who can use it: Users with access to Grok Build’s plugin marketplace; the Expo plugin can be installed or updated from the catalog.

    How to use it: Run the `/marketplace` command in Grok Build and select Expo. For a new installation, list the plugins with `grok plugin marketplace list` and run `grok plugin install expo --trust`.

    Tip: Try the update in an Expo development or preview environment before using it in a live store release.

  • Plugin / IntegrationGrok Build

    Stripe Grok Build plugin updated to version 0.9.47

    The pinned version of the Stripe plugin in Grok Build’s official plugin marketplace was upgraded from 0.9.46 to 0.9.47. The plugin directory in the catalog and the commit pin were also updated; the official commit does not provide a detailed feature list.

    Usage details

    Who can use it: Users with access to Grok Build’s plugin marketplace; the Stripe plugin can be installed or updated from the catalog.

    How to use it: Run the `/marketplace` command in Grok Build and select Stripe. For a new installation, list the plugins with `grok plugin marketplace list` and run `grok plugin install stripe --trust`.

    Tip: Use Stripe test environment keys and the permissions requested by the plugin before running payment or billing operations.

  • Plugin / IntegrationGrok Build

    Vercel Grok Build plugin updated to version 0.55.1

    The pinned version of the Vercel plugin in Grok Build’s official plugin marketplace was upgraded from 0.55.0 to 0.55.1. The plugin directory in the catalog and the commit pin were also updated; the official commit does not provide a detailed feature list.

    Usage details

    Who can use it: Users with access to Grok Build’s plugin marketplace; the Vercel plugin can be installed or updated from the catalog.

    How to use it: Run the `/marketplace` command in Grok Build and select Vercel. For a new installation, list the plugins with `grok plugin marketplace list` and run `grok plugin install vercel --trust`.

    Tip: After updating, verify the Vercel workflows you use, such as deployments and log access, in a small test project.

  • Plugin / IntegrationGrok Build

    Wix Grok Build plugin updated to version 1.41.4

    The pinned version of the Wix plugin in Grok Build’s official plugin marketplace was upgraded from 1.41.3 to 1.41.4. The plugin directory in the catalog and the commit pin were also updated; the official commit does not provide a detailed feature list.

    Usage details

    Who can use it: Users with access to Grok Build’s plugin marketplace; the Wix plugin can be installed or updated from the catalog.

    How to use it: Run the `/marketplace` command in Grok Build and select Wix. For a new installation, list the plugins with `grok plugin marketplace list` and run `grok plugin install wix --trust`.

    Tip: Before updating, check that the Wix account, site, and MCP permissions are linked to the correct workspace.

Tencent · Hy

Tencent Hy4
Full history (369) →

New in this edition · 28

  • APIteamai-cli

    --replace-token added to replace an HTTP agent token without checking the old value

    init --http and source add-http can now replace a working sk- agent token with a new one using --replace-token, without trying the old token. Other agent tokens specified with --agent and workspace bindings are preserved; this option is not needed for the shared hk- key.

    Usage details

    Who can use it: teamai-cli HTTP local-agent setup and HTTP source configuration.

    How to use it: Run teamai init --http https://your-team-host/api --agent workbuddy --token <new-key> --force --replace-token or, for an existing endpoint, teamai source add-http <endpoint> --agent workbuddy --token <new-key> --replace-token.

    Tip: Because the new token may invalidate the old one on the backend, resync the model configuration after running the command.

    ✓ Verified by official sourceSource: teamai --replace-token commit’i
  • ModelHy Realtime

    Hy Realtime adds search and singing to live voice conversations

    Tencent Hunyuan released Hy Realtime; the model infers intent from unfinished speech, searches the web when needed, and can sing on request. IT之家 and Pandaily report that the model is now available in Yuanbao voice conversations and Tencent customer service, while the Tencent Cloud API is coming soon.

    Usage details

    Who can use it: Using the Chuyu/初语 voice in voice conversations in the latest Yuanbao app; also available for up to 10 minutes of testing in Tencent Cloud’s Hy Realtime web demo. The Yuanbao and API rollout is reported by secondary sources.

    How to use it: Update Yuanbao, start a voice conversation, and select the Chuyu/初语 voice; in the web demo, start speaking and enable online search.

    Tip: Enable online search when you need up-to-date information; the demo notes that this option may slow responses.

  • AvailabilityLoopForge

    Conversation roots moved to AGENTFLOW_CONVERSATION_DIR, accessible to Docker

    Local conversation roots are now stored under AGENTFLOW_CONVERSATION_DIR rather than /data/assistant, which the Docker daemon cannot access. Compose sets this to LOOPFORGE_DATA_DIR/conversations, fixing workspace startup in Docker Desktop and similar setups.

    Usage details

    Who can use it: LoopForge platform installations running with Compose; users with a customized compose.yaml should also add the variable.

    How to use it: Set AGENTFLOW_CONVERSATION_DIR to ${LOOPFORGE_DATA_DIR}/conversations in compose.yaml, update the checkout, and run ./start.sh pull.

    Tip: If you have a custom compose file, don’t assume that old conversation roots under /data/assistant were moved automatically; back up your data and verify the new directory.

  • CapabilityLoopForge

    Sandboxes with custom images can automatically use the platform toolchain

    Custom Dockerfile sandboxes that don’t include Node.js, npm, or Python 3 can now get the platform’s Node.js, npm, portable Python 3.12, and bundled CLIs on first use. The toolchain is mounted read-only into the workspace and added to the end of PATH; images using musl or a foreign architecture fall back to their own package-manager flow.

    Usage details

    Who can use it: LoopForge conversation workspaces using custom Docker images; the built-in platform image does not use this fallback path.

    How to use it: Open a conversation workspace with a custom Dockerfile; when a native CLI is needed, LoopForge prepares the toolchain under AGENTFLOW_RUNTIME_CLI_DIR/toolchain/<build> and runs the command automatically.

    Tip: Each platform build uses about 200 MB, plus space for additional CLIs; after upgrading, remove unused old toolchain/<build> directories.

  • Settingteamai-cli

    Role requirement setting added for selected skill namespaces

    sharing.skills.roleRequiredNamespaces in teamai.yaml ensures that selected skill namespaces are delivered only to members whose effective role includes that namespace. Members without the role skip these namespaces during pull; tags and workspace child selections are subject to the same filter.

    Usage details

    Who can use it: teamai-cli team repositories; all members need to upgrade to a version that supports this setting.

    How to use it: Add sharing.skills.roleRequiredNamespaces: [bus_common, ebao, health] to teamai.yaml; on the member side, run teamai roles list, teamai roles set developer, and teamai pull.

    Tip: Enter names as a single directory segment; do not use paths or globs, and test delivery with a small namespace first.

  • CapabilityLoopForge

    SOP logs redact agent keys and secret environment variables

    Run logs and telemetry now mask environment variables ending in _API_KEY, _TOKEN, _SECRET, or _PASSWORD. Values such as DEEPSEEK_API_KEY and AGENTFLOW_CLI_API_KEY used with a custom Base URL should no longer appear in clear text in the docker exec line.

    Usage details

    Who can use it: LoopForge SOP runs and platform telemetry; keys already written to logs before upgrading are unaffected.

    How to use it: Upgrade LoopForge, run a new SOP, and verify that values are redacted in the logs; if an SOP ran on an older version, rotate the relevant key.

    Tip: In addition to known suffixes, audit your custom secret variables for log access.

  • APIteamai-cli

    teamai push --file and --path options are back

    The teamai push --file <local file> --path <team path> workflow, accidentally removed in 0.30.0, has been restored. The command proposes only the specified file at the specified team path as a pull request and does not carry over other changes in the team clone.

    Usage details

    Who can use it: Members using direct files and scripts after teamai-cli 0.30.0.

    How to use it: Run teamai push --file <local file> --path <team path>; if the branch contains the default file bytes, the command also reports that there are no changes to push.

    Tip: For single-file changes, use --file/--path to avoid accidentally including other edits from the same clone in the PR.

    ✓ Verified by official sourceSource: teamai push --file/--path commit’i
  • CapabilityLoopForge

    Upgrade flow reapplies missing schema extensions

    ./start.sh pull now reapplies the run_telemetry.adapter and run_logs.level generated columns, which GORM AutoMigrate cannot create, as well as the initial catalog entries. Old data volumes should no longer fail on startup with an Unknown column adapter error.

    Usage details

    Who can use it: LoopForge upgrades; the installer and upgrade use the same scripts/install-schema-extensions.sh flow.

    How to use it: Update the checkout and run ./start.sh pull; if the step fails, manually apply the migration or scripts/install-schema-extensions.sh file indicated in the output.

    Tip: Back up the database before a production upgrade, then open the run detail page to verify the generated columns.

  • CapabilityLoopForge

    CodeBuddy’s built-in plugin marketplaces are no longer downloaded for every conversation

    LoopForge automatically sets CODEBUDDY_SKIP_BUILTIN_MARKETPLACE=1 in CodeBuddy processes, preventing the two built-in marketplaces from being downloaded for every conversation and workspace. Usage per new conversation drops from about 72 MB to under 1 MB; LoopForge’s own skill and agent delivery is unaffected.

    Usage details

    Who can use it: LoopForge platform conversations and workspaces; marketplaces that were previously downloaded continue to work in existing conversations.

    How to use it: Upgrade LoopForge; the platform automatically adds the setting to CodeBuddy processes. If you start a CodeBuddy process outside LoopForge, set the CODEBUDDY_SKIP_BUILTIN_MARKETPLACE=1 environment variable yourself.

    Tip: If you use your own plugin marketplace, test the effect of automatic skipping; LoopForge does not install plugins in this flow.

  • CapabilityLoopForge

    Codex models are no longer accidentally replaced in CodeBuddy-only configurations

    A routing bug that caused LoopForge to replace the Codex model with the CodeBuddy model in CodeBuddy-only entries has been fixed. This preserves model selection in workspaces that use different agent CLIs with the same configuration.

    Usage details

    Who can use it: CodeBuddy-specific model configurations in LoopForge chats and SOPs.

    How to use it: Select the CodeBuddy-specific model in personal settings or agent configuration, reopen the workspace, and check that the model is preserved in the CodeBuddy session.

    Tip: In projects that use multiple CLIs, specify the model explicitly for each agent instead of relying on the global default.

  • Capabilityteamai-cli

    Non-git workspace subdirectories find the nearest teamai configuration

    A session started from a subdirectory of a non-git workspace now finds the nearest .teamai/config.yaml or .teamai/teamai.yaml file in parent directories and binds delivery to that root. This prevents the subdirectory from accidentally falling into another user scope or the general scope.

    Usage details

    Who can use it: Non-git folder workspaces and their subdirectories.

    How to use it: Create the configuration with teamai init or teamai bind-project in the root folder, then cd into the subdirectory and run teamai pull or start an AI session.

    Tip: To avoid the wrong scope in nested workspaces, don’t inadvertently add a second .teamai configuration to subfolders.

    ✓ Verified by official sourceSource: teamai non-git workspace commit’i
  • Capabilityteamai-cli

    Local-agent report shows prompt and hook deliveries with their types

    The local-agent report now distinguishes rule files read from disk, prompts delivered through claudemd, and agent hooks using rule_type fields. The report counts only inputs actually received by the current tool and scope; it marks items whose source is absent from the manifest or whose version is unknown as local or unversioned.

    Usage details

    Who can use it: teamai-cli HTTP local-agent reports; depends on the tool and user/workspace scope.

    How to use it: Upgrade teamai-cli and start a new AI session in the workspace connected to the HTTP agent; inspect the rule_type=rule, prompt, or hook fields in the next local-agent report.

    Tip: When interpreting Enterprise counts, remember that the same slug can overlap as a file, prompt, and hook; use the report’s rule_type value.

    ✓ Verified by official sourceSource: teamai local-agent rapor commit’i
  • Plugin / Integrationteamai-cli

    MCP git-exclude checks no longer raise false warnings for targets outside the project

    Project-scope MCP git-exclude checks no longer treat MCP targets in the user’s HOME or in non-git directories as project files. This removes false unknown tracking warnings and unnecessary exclude checks for non-project targets.

    Usage details

    Who can use it: teamai-cli project scope, MCP reconciliation, and teamai doctor checks.

    How to use it: Upgrade teamai-cli and run teamai doctor or the MCP pull/reconcile flow in project scope; only actual project targets are checked for git-exclude.

    Tip: Leave personal MCP configurations in place under the user HOME scope instead of copying them into the project repository.

  • Plugin / IntegrationLoopForge

    The platform image brings DeepSeek Harness and more CLIs to the SOP sandbox

    The platform image bundles @deepseek-ai/dsh@0.2.2-rc.1 alongside Codex and CodeBuddy; SOP sandboxes can also install agent CLIs on first use. This lets users with a DeepSeek key run conversation and SOP roles without additional setup.

    Usage details

    Who can use it: LoopForge platform image 0.3.8 and later; DeepSeek Harness uses Node 24.9 and a pinned package version.

    How to use it: Upgrade the platform image, select DeepSeek Harness in a conversation or SOP role, and configure DEEPSEEK_API_KEY; if the image isn’t being used, wait for the pinned DSH package to be installed for the first time.

    Tip: Allow for an approximately 180 MB increase in image download size, and don’t change the Node version required by the DSH prerelease.

  • APIteamai-cli

    Self-mode push now accepts new and remote-only branch targets

    Self-mode teamai push --branch now accepts new branches and remote-only branch targets, expanding the flow that previously worked only with suitable existing targets. Subsequent pushes to the same open branch add new changes while preserving previous files.

    Usage details

    Who can use it: teamai-cli self-mode; open-branch flows where the local or remote branch is not yet visible in the local clone.

    How to use it: Run teamai push --all --branch release/knowledge; a new branch is created, and you can reuse the same command for subsequent changes.

    Tip: Before targeting an existing branch with someone else’s unpushed commit, heed teamai’s warning; choose a new branch as the safe target.

  • CapabilityLoopForge

    Shared native roots are also cleaned up when conversations are deleted

    When a conversation is deleted, its shared native root is no longer left on disk; the daily janitor also cleans up roots left by older versions that have not changed for a day. This limits disk growth for tasks that create a new conversation on every run.

    Usage details

    Who can use it: LoopForge platform conversations and the daily cleanup task.

    How to use it: Delete the conversation through the platform interface; existing old roots are cleaned up during a janitor run after the conversation has been removed and the roots have remained unchanged for a day.

    Tip: On large, older installations, monitor the size of the conversations directory after upgrading; don’t manually delete active conversation roots.

  • Capabilityteamai-cli

    30-second request timeout added to TGit organization imports

    During TGit --from-org imports, groups/projects pages and the fallback endpoint are now protected by a 30-second AbortSignal.timeout, including the response body. Stalled requests fail instead of returning a partial repository list.

    Usage details

    Who can use it: teamai-cli TGit/Gongfeng organization import workflow.

    How to use it: Run teamai import --from-org <group> as usual; requests that exceed 30 seconds return an error and do not continue with an incomplete whitelist.

    Tip: If you time out in a large organization, fix the network or TGit API access and restart the import; do not treat a partial list as reliable.

  • AvailabilityLoopForge

    openssh-client restored to the built-in generic-backend image

    The generic-backend Dockerfile now includes openssh-client to match the platform image. This lets generic-docker workspaces or SOP sandboxes start with the platform image on first use instead of building a separate agentflow-img image.

    Usage details

    Who can use it: LoopForge generic-docker workspaces and SOP sandboxes; existing installations may require the 031_generic_backend_openssh migration.

    How to use it: Update the checkout and run ./start.sh pull; apply the required migrations/031_generic_backend_openssh.sql step to the old data volume through the installer or upgrade flow.

    Tip: Before manually changing the custom generic-backend entry, check whether the current catalog value has been edited by a user.

  • FeatureLoopForge

    Artifact preview and source selector resized to match the toolbar

    The artifact preview/source toggle in the chatbot interface has been resized to match the other toolbar buttons. Its functionality is unchanged, but preview and source views now offer more consistent controls.

    Usage details

    Who can use it: LoopForge chatbot web interface.

    How to use it: Open an artifact and use the preview/source toggle; no separate setting or installation is required.

    Tip: Use source view for debugging and preview for a quick visual check.

    ✓ Verified by official sourceSource: LoopForge changelog commit geçmişi
  • Featureteamai-cli

    teamai hook types are labeled in the Codex /hooks view

    Codex’s /hooks review list shows team hooks as ID, HTTP agent hooks as slug, built-in hooks as teamai-<Event>, and shared dispatchers as teamai-team-<Event>. This makes it easier to distinguish a hook’s source in the review UI.

    Usage details

    Who can use it: teamai-cli’s Codex hook setup and the Codex /hooks review interface.

    How to use it: Upgrade teamai-cli, open the /hooks list in Codex, and check the relevant teamai hook label.

    Tip: Check the label type before removing a hook; an ID or slug may differ from a user hook for the same event.

  • Plugin / IntegrationLoopForge

    Token snippets copied from HTTPS Studio no longer generate HTTP

    In Studio accessed over HTTPS, copied MCP, REST, and quick-access snippets no longer generate token-bearing http:// endpoints; the snippets use https://.

    Usage details

    Who can use it: LoopForge v0.3.8 and HTTPS Studio connections.

    How to use it: Open Studio over HTTPS and copy the relevant MCP, REST, or quick-access snippet; replace old http:// addresses in your client with the new snippet.

    Tip: Check copied configurations for an http:// and Bearer combination using grep or a similar search.

    ✓ Verified by official sourceSource: LoopForge v0.3.8 sürüm notları
  • CapabilityLoopForge

    LoopForge strips credentials from repository URLs

    LoopForge now rejects saving cloneUrl/webUrl values containing a username or password; for existing records, it strips credentials from API, workspace, and runner views. Git clone errors and redacted Git output will also mask this information.

    Usage details

    Who can use it: LoopForge platform repositories and workspaces; previously saved URLs may remain in the old format until they are saved again.

    How to use it: First save the Git token in personal settings, then save the repository in Studio again using a URL without a token; do not use private URLs in the https://user:token@... format.

    Tip: Resave old URLs immediately after upgrading, and rotate any tokens that may have been written to logs previously.

  • SettingLoopForge

    Identity headers behind a proxy are no longer trusted without a secret key

    When AGENTFLOW_TRUST_IDENTITY_HEADERS=true, identity headers such as Staffname are no longer accepted directly on requests carrying X-Forwarded-For, Forwarded, or X-Real-IP. The gateway must send a secret matching AGENTFLOW_IDENTITY_HEADER_SECRET in the X-AgentFlow-Gateway-Secret header.

    Usage details

    Who can use it: LoopForge installations behind a reverse proxy that use identity-header login; installations with the setting disabled are unaffected.

    How to use it: Set AGENTFLOW_IDENTITY_HEADER_SECRET on the gateway and add the matching X-AgentFlow-Gateway-Secret header; test login through the proxy after upgrading LoopForge.

    Tip: Do not expose the secret key to clients outside the proxy; if you are not using this setting, leave AGENTFLOW_TRUST_IDENTITY_HEADERS disabled.

  • Capabilityteamai-cli

    30-second request timeout added to GitCode organization imports

    GitCode --from-org repository-listing calls now use AbortSignal.timeout with a 30-second limit, including the response body. A stalled page stops the import; no whitelist is written with an incomplete or partial repository list.

    Usage details

    Who can use it: teamai-cli GitCode organization import flow.

    How to use it: Run teamai import --from-org <org>; if you get a timeout, fix GitCode access and then rerun the command.

    Tip: When an import fails, verify access to the source organization and the token instead of manually using partial results.

  • AvailabilityLoopForge

    MIRROR registry option added for LoopForge platform images

    If GHCR is slow, start.sh can now use the MIRROR variable to pull a pinned platform image from another registry. The downloaded image can still be checked against the pinned name and digest in GHCR.

    Usage details

    Who can use it: LoopForge v0.3.8 platform installations; example mirror: ghcr.nju.edu.cn.

    How to use it: Run MIRROR=<host> ./start.sh or, for an existing installation, MIRROR=<host> ./start.sh pull; then check the digest with docker image inspect.

    Tip: Set MIRROR only to a registry you trust; do not deploy to production without verifying against the digest in the release notes.

  • CapabilityAI-Infra-Guard

    MCP-Scan no longer drops findings when optional tags are missing

    If a finding lacks level or suggestion tags, MCP-Scan now preserves those fields with empty values instead of silently dropping the entire finding. Findings are still skipped if the required title, desc, or risk_type fields are missing.

    Usage details

    Who can use it: MCP-Scan use with AI-Infra-Guard v4.6.10 and later.

    How to use it: Upgrade AI-Infra-Guard to v4.6.10 and run MCP-Scan with your usual command; check the results list even if the model response does not include level or suggestion.

    Tip: After scanning, fill empty level/suggestion fields in reports using your default severity and recommendation policies.

  • CapabilityLoopForge

    Sandbox Git cleanup preserves index.lock on a live alternate path

    Sandbox Git cleanup does not delete an index.lock file while a process is still using that same file through an alternate bind-mounted or hard-linked path. This prevents ongoing Git operations from being disrupted during upgrades or cleanup.

    Usage details

    Who can use it: LoopForge v0.3.8 sandbox Git operations.

    How to use it: Upgrade to LoopForge v0.3.8; no special activation is required, and processes holding locks are automatically preserved during workspace cleanup.

    Tip: When you encounter a lock error, terminate the running Git or sandbox process first instead of deleting index.lock manually.

    ✓ Verified by official sourceSource: LoopForge v0.3.8 sürüm notları
  • APIteamai-cli

    teamai.yaml changes are merged on refresh even without selecting a branch

    In self-mode, local changes to teamai.yaml are now merged with new settings on the default branch during the regular teamai push refresh, not only with --branch. If the same setting has changed differently on both sides, push reports a conflict and does not open a pull request.

    Usage details

    Who can use it: teamai-cli self-mode team repositories.

    How to use it: Edit teamai.yaml and run teamai push --all; if there is an incompatible key, fix the output and push again.

    Tip: If you change team settings frequently, reduce the scope for conflicts with small, single-purpose teamai.yaml commits.

Watchlist

Kimi K3.1 Pro

46
Moonshot AI · Kimi

In 11th place; still the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. Kimi Code 2.6.0 reached stable status; Kimi Work 3.2.27 can run long tasks in the background.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; 12th place 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 about 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 as `muse-ember-1.0` in Meta Model API and on all Muse Code plans.

Grok 4.7

35
xAI · Grok

The current model for the Grok app and API. Grok 4.8 is expected to become available to some Grok Bot users; its official API ID, price, and Artificial Analysis score are not yet available. Grok Bot 1.0 was released on November 11.

Tencent Hy4

35
Tencent · Hy

Available through the API on OpenRouter (`tencent/hy4`); 50% off on TokenHub through December 31, and free on WorkBuddy and CodeBuddy through November 15. The older Hy-Image-3.0 image model was removed from TokenHub on November 12.

Sources