YZ·RADARyz.mikail.net
Daily edition · 20 November 2026

TOP 10 holds steady; computer use arrives in Gemini’s Mac app

There are no changes in the TOP 10 today; both rankings and scores in Artificial Analysis Intelligence Index v5.0.1 are unchanged from yesterday. Claude Fable 5.2 (65) leads, followed by GPT-6.2 Astra (62), Gemini 4 Neon (61), Gemini 4 Carbon (60), Claude Opus 5.5 (59), GPT-6.2 Sol (57), Claude Sonnet 5.5 (55), Gemini 4 Hydrogen (54), GPT-6.1 Terra (51) and Gemini 4 Argon (50). Google has four models on the list, while Anthropic and OpenAI have three each. Muse Wonder Full is tied with Argon at 50 points but remains out of the TOP 10 because it ranks 11th in the table. The watchlist is also unchanged: Muse Wonder Full (50), Kimi K3.1 Pro (46, the strongest open-weight model), Muse Wonder (44), Qwen4 Max (43), Claude Haiku 5.5 and Muse Ember (36). Elon Musk wrote that Grok 4.8 is being tested not as a single model but in several beta variants, and that it is not yet clear which one will be released for wider use; Grok 4.8 is not yet in the Artificial Analysis table. LMArena’s live page was again blocked by a CAPTCHA; the latest accessible view is dated November 11, with Gemini 4 Carbon, Gemini 4 Argon and Muse Wonder in the top three.

Google’s most visible launch of the day is Computer use for Gemini app for Mac. Gemini can now carry out multistep tasks in desktop apps and websites in a separate window, clicking and typing to do things like fill out forms. Users choose which apps to allow, watch a live preview and approve sensitive steps such as signing in or making payments. For now, it is available only in the U.S., in English, on Apple Silicon Macs (macOS 15 and later), for personal accounts belonging to people over 18 who subscribe to Google AI Plus, Pro or Ultra. Gemini can also import a preference summary and chat history from a ZIP archive exported from another AI assistant. For developers, the Gemini API’s Interactions API now has a Tool Retrieval preview: instead of sending the schemas for large tool libraries with every request, tools are searched and loaded when needed, reducing input tokens and latency. Genie 4 is also available to developers in the paid tier of the Interactions API and AI Studio; it generates interactive worlds from text or images at 24 fps and 720p. Gemini 4 Neon is generally available in Gemini Enterprise Agent Platform, Gemini Enterprise Workflow Builder and BigQuery generative AI features. In BigQuery, natural-language search using `CREATE AI INDEX` has entered preview, without requiring embedding management. CodeMender supports Gemini 4 Hydrogen in preview, while Gemini CLI 0.70.0-preview.0 has tightened MCP security. At OpenAI, Codex 0.171.0 added GPT-6.2 Sol and Astra to the model catalog and set Sol as the default. The TUI workspace view is now the default, with previews for documents, images and subagents opening alongside it. `codex exec --remote <ADDR>` lets automations connect to a running app server session, while `/rc` and `/remote` commands manage remote control and device pairing. In addition, the `codex sandbox` command runs any command in a sandbox using a named permissions profile in config.toml. ChatGPT Health has expanded in the U.S. to Android Health Connect, the desktop app, dots and Sites for Free, Go, Plus and Pro users. Dots has begun testing with a small number of Plus users and rolling out to all Business Standard users. An Accounting app for accounting tasks has arrived in the desktop app. The Python SDK has added the `gpt-live-1.1` model and standby support for Live WebRTC sessions.

All 36 of Anthropic’s entries come from Claude Code v2.1.329, released on the night of November 19; there were no new model or Claude app announcements over the same period. `/tasks` now shows a session’s scheduled jobs, `/loop` wake-ups and conversation forks in one list; it asks for confirmation before stopping a task with `x`. One change to note: most scheduled jobs saved by older versions must be rescheduled in newly logged-in sessions. Code Mode now works with OpenTelemetry events and metrics enabled. In Code Mode, a failing PreToolUse hook no longer blocks a script call; hooks that should block it must include `"onFailure": "block"`. On macOS, computer use now follows stricter security rules and gives up on apps that do not open within 30 seconds. If no target is specified, `/code-review` reviews the branch against the default branch and reads PRs using `gh pr diff`. Auto mode runs `git status` less often in large repositories. The `max_tokens` automatic retry introduced in 2.1.326 has been removed. Fixes address `~/.claude.json` being reset during network lockups and PDFs, images and Office files from MCP being saved as `.bin`. Moonshot’s 40 entries come from Kimi Code. Kimi Code CLI 2.7.0 can connect to Amazon Bedrock models using `type = "bedrock"` and AWS’s standard credential chain. Selecting an SSH Host runs the full Kimi Code server on a remote machine, so the remote session continues even if the local computer sleeps. The KAP v1 protocol has been merged, and `kimi web --json` and a `:wait` endpoint have been added so another agent can monitor Kimi Code over HTTP only. REST API v2 routes have been removed; older integrations may receive 404 errors. Kimi Code Desktop 1.4.0 can connect to and take over Claude Code, Codex and Cursor sessions, adds Qwen Coding Plan as a provider, and now uses messages sent during an active turn to steer it immediately by default.

Of Alibaba’s 57 entries, 40 come from Qwen Code v0.28.0-nightly: a GitHub adapter for managed channels, `/send` for communication between sessions, general-purpose external webhook delivery, a Traditional Chinese interface and idle-session suspension in Hosted Harness. Qoder 0.5.16 adds remote access to devices on the same account across different networks (once Allow control is enabled on the target), the ability to assign Issues to Agent Teams, editing password-protected Word and Excel files while preserving their encryption, sharing conversations and output files, PAC proxy support and output-limit settings for custom models. QoderWake 1.2.20 includes a Catch up missed runs option that starts missed scheduled runs after a device wakes from sleep. Tencent’s 28 entries come from open-source tools: AI-Infra-Guard’s Prompt-Security scanner now has a multiround RefusalLadderJailbreaking operator. BrowserSkill is more stable on single-page apps and heavyweight web apps such as Lark Base. LoopForge installs its Linux helper automatically in macOS Docker workspaces, while WeKnora shows cloning progress in Lite mode without Redis. At xAI, personal MCP servers saved to Grok Bot are now subject to a team allowlist and a limit of 100 servers; photos and PDFs sent from a phone reach the Bot’s computer. xAI TypeScript SDK v0.2.9 fixes a stream-parsing error. At Meta, WhatsApp Cloud API added a call quality score with GREEN, YELLOW and RED values for business calls and the `CALLING_QUALITY_UPDATE` webhook event (currently for numbers in Colombia and Mexico); the Forum app for Facebook Groups has launched on Android in Indonesia. Xiaomi has added MiMo-V2.6-Pro-UltraSpeed support to Token Plan Pro and Max for a limited time and updated the V2.6 and Batch API pricing table. Z.ai has one unverified report: according to a user notification, ZCode launched a 10-hour campaign offering a free 300M GLM-5.3-Flash quota on the night of November 18–19; the campaign has ended. DeepSeek had no new user-facing changes between November 18 and 20. A total of 218 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; leads with 65 points, with its score and ranking unchanged. Available on all paid Claude plans; `claude-fable-5-2` in the API, $8 / $40, with thinking always adaptive. The Claude Meetings beta is available on Team and Enterprise plans (Google Meet first).Released: 9 November 2026

  2. 02

    GPT-6.2 Astra

    OpenAI · ChatGPT=
    62pts

    Released on November 17; remains in second place with 62 points. Costs $4,00 per task at Max. `gpt-6.2-astra` in the API, $8 / $40, cached input $0,40; 1.05M context, 128K output. Codex 0.171.0 added Astra to the model catalog, but Sol is the default; Astra must be selected manually in the model selector for harder tasks. Also available in GitHub Copilot (Pro+, Max, Business, Enterprise).Released: 17 November 2026

  3. 03

    Gemini 4 Neon

    Google · Gemini=
    61pts

    Released on November 18; remains in third place with 61 points. Costs $6,44 per task at xhigh. `gemini-4-neon` in the API (paid tier only), $3 / $15, 1M context; low, medium, high and xhigh thinking levels. Also generally available in Gemini Enterprise Agent Platform: supports a combined input/output limit of 1,048,576 tokens, `continuation_token` for long outputs and an OpenAI-compatible Chat Completions API. Also selectable in Gemini Enterprise Workflow Builder and BigQuery generative AI features.Released: 18 November 2026

  4. 04

    Gemini 4 Carbon

    Google · Gemini=
    60pts

    Score (60) and ranking unchanged. `gemini-4-carbon` in the Gemini API and AI Studio (paid tier only; $2 / $10 through November 30, then $4 / $20 starting December 1). Carbon runs the security loop in the Stitch Loop agent alongside CodeMender.Released: 28 October 2026

  5. 05

    Claude Opus 5.5

    Anthropic · Claude=
    59pts

    Score (59) and ranking unchanged. Anthropic’s flagship Opus model; available in Claude Pro/Max/Team/Enterprise and the API, as well as through Bedrock, Google Cloud and Azure.Released: 22 September 2026

  6. 06

    GPT-6.2 Sol

    OpenAI · ChatGPT=
    57pts

    Released on November 17; remains in sixth place with 57 points. Costs $1,01 per task at Max, one-quarter the price of Astra. `gpt-6.2-sol` in the API, $2 / $10, cached input $0,10; 1.05M context, 128K output. It became the default model in the Codex model catalog with Codex 0.171.0. Also available in ChatGPT Work and GitHub Copilot.Released: 17 November 2026

  7. 07

    Claude Sonnet 5.5

    Anthropic · Claude=
    55pts

    The strongest model available on claude.ai’s free plan; its score (55) and ranking are unchanged.Released: 28 September 2026

  8. 08

    Gemini 4 Hydrogen

    Google · Gemini=
    54pts

    Released on November 12; remains in eighth place with 54 points. Costs $3,16 per task at xhigh. `gemini-4-hydrogen` is also available in the free API tier; paid pricing is $1,50 / $7,50. Tasks runs on Gemini Hydrogen. Google’s security vulnerability discovery and remediation tool, CodeMender, has supported Hydrogen in preview since November 19; the default model is still Gemini 3.8 Flash.Released: 12 November 2026

  9. 09

    GPT-6.1 Terra

    OpenAI · ChatGPT=
    51pts

    Released on October 26 for ChatGPT Work, Codex and the API; focused on vision and computer use. Score (51) and ranking unchanged. There is no GPT-6.2 version yet.Released: 26 October 2026

  10. 10

    Gemini 4 Argon

    Google · Gemini=
    50pts

    Available to Google AI Pro/Ultra subscribers and through the paid API; introductory price of $2 / $10 through November 30. Tied with Muse Wonder Full at 50 points; it holds tenth place because it ranks ahead of Muse Wonder Full in the Artificial Analysis table.Released: 30 September 2026

Key signals

  • The TOP 10 is unchanged: Claude Fable 5.2 (65) leads, Gemini 4 Neon (61) is third, and all scores are the same.
  • Computer use has arrived in Gemini app for Mac; Tool Retrieval preview and Genie 4 are now available in the Gemini API.
  • Codex 0.171.0: GPT-6.2 Sol is the default model; TUI workspace and `codex exec --remote` have arrived.
  • Claude Code 2.1.329: unified `/tasks`, OpenTelemetry in Code Mode, and stricter security for macOS computer use.
  • Kimi Code adds Amazon Bedrock and an SSH server; Qoder 0.5.16 adds device control across networks.
  • 218 new developments: Alibaba 57, Moonshot 40, Anthropic 36, Tencent 28, OpenAI 20, Google 16, xAI 15, Meta 3, Xiaomi 2, Z.ai 1.

Worth trying today

01

Load large tool sets in the Gemini API with Tool Retrieval

If your agent sends dozens of tool definitions, including all the schemas with every request adds token and latency costs. In the Interactions API, add the `{"type": "tool_search"}` tool to the request and set `defer_loading: true` for tools that are not needed in every request; the model will search for and load the schemas it needs. If you manage your own tool registry, use `execution: "client"` to search on the client side. The feature is in preview; it works with Gemini 3.6 and later and all Gemini 4 models, and remote MCP servers can also use deferred loading. First, run the same task in a test project with Tool Retrieval both on and off, then compare the input token counts.

Google · Gemini
02

Use Sol for everyday Codex work and Astra for harder tasks

Update Codex to 0.171.0; GPT-6.2 Sol is the default in the model catalog. According to Artificial Analysis, Sol costs about $1,01 per task, while Astra costs $4,00; leave routine editing, testing and explanation tasks to Sol, and switch to GPT-6.2 Astra in the model selector for architectural decisions and difficult debugging. For automations, connect to a running app server session with `codex exec --remote <ADDR>` instead of opening a new session every time. To control Codex from another device, enter `/rc` or `/remote` in the TUI and use the pairing code. To run a tool with Codex’s sandbox restrictions, define a permissions profile in config.toml and use `codex sandbox` with that profile.

OpenAI · ChatGPT
03

Check scheduled jobs and hooks after Claude Code 2.1.329

After updating to 2.1.329 with `claude update`, open `/tasks`; scheduled jobs, `/loop` wake-ups and forks are now listed there. Most scheduled jobs saved in an older version will not run in a newly logged-in session; reschedule any that do not appear in the list. If you use Code Mode, a failing PreToolUse hook no longer stops a script call; add `"onFailure": "block"` to command, HTTP or MCP hooks that must block it. If you are behind a gateway with a lower output limit, set `CLAUDE_CODE_MAX_OUTPUT_TOKENS` to that limit because automatic retries have been removed. If you run a gateway, make sure `access_control.allow_cidrs` is not empty; enter the allowed IP ranges.

Anthropic · Claude
04

Run Kimi Code with Bedrock or a remote SSH server

Update to Kimi Code 2.7.0. If you use AWS, add `type = "bedrock"` under `[providers.bedrock]` in `config.toml`, set the `AWS_REGION` variable and sign in with AWS SSO, a shared profile or an instance role; a long-lived Bedrock key is not required. If needed, add `protocol = "anthropic"`, `"openai"` or `"openai_responses"` to the model alias. For long-running jobs, select an SSH machine when creating a project and also sign in to your Kimi account on that machine; the job continues on the remote server even if your laptop sleeps. Use `kimi server` for a headless server, or `kimi server --open` to open the browser too. With Kimi Code Desktop 1.4.0, you can also take over Claude Code, Codex or Cursor sessions in the same interface.

Moonshot AI · Kimi

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#5 Claude Opus 5.5#7 Claude Sonnet 5.5Claude Haiku 5.5
Full history (836) →

New in this edition · 36

  • FeatureClaude Code

    `/code-review` reviews against the default target branch and verifies the PR diff

    For most models, `/code-review` without a specified target reviews the current branch against the default branch. When given a PR number with `--fix` or a PR URL, it stops if `gh pr diff` fails or is rejected rather than performing an incorrect review.

    Usage details

    Who can use it: Claude Code 2.1.329; most models and repositories with GitHub CLI access.

    How to use it: Use `/code-review` for a branch review; for a PR, use `/code-review <pr-number> --fix` or a PR URL.

    Tip: Before reviewing a PR, test that `gh pr diff` works with the same credentials.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • FeatureClaude Code

    `/tasks` now brings together scheduled jobs, `/loop` wakeups, and forks

    The `/tasks` view also lists the current session's scheduled tasks, `/loop` wakeups, their targets, and conversation forks. Pressing `x` to stop an item now prompts for confirmation first.

    Usage details

    Who can use it: Sessions using Claude Code 2.1.329.

    How to use it: Type `/tasks` in Claude Code; select the item you want to stop, press `x`, then choose the confirmation option.

    Tip: Use this view as a single checkpoint instead of checking background job statuses separately.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AgentClaude Code

    Subagent stopping, wake-ups, and model/effort handoff are more consistent

    Fixed stopped background agents restarting during resume, and completed subagents inheriting their sibling's model/effort settings. Tasks completed in the background are reported to the main conversation more accurately; the Agent tool's issue ignoring the `effort` parameter for named teammates was also fixed.

    Usage details

    Who can use it: Claude Code 2.1.329; background agents and agent teams.

    How to use it: Run a background agent or agent team and use the stop/resume flow; check the completion notification in the main conversation.

    Tip: Set the model and effort values explicitly for named teammates.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code

    Settings persistence and macOS session continuity fixed for network lockups

    An unreadable `~/.claude.json` during a network outage or file lock is no longer reset to defaults. On macOS, if Keychain cannot be read or written, the relevant session now reports the issue until it expires and prevents other sessions from signing out unnecessarily.

    Usage details

    Who can use it: Claude Code 2.1.329; settings file on all platforms, Keychain on macOS.

    How to use it: Update to this version and use settings and session operations as usual; if the issue persists, check `~/.claude.json` and Keychain permissions.

    Tip: Avoid having tools that update the settings file conflict with Claude Code over file locks.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Code Mode hook, memory, and script constraints are clearer

    In Code Mode, a failing PreToolUse hook no longer rejects the script tool by default; `onFailure: "block"` is required to reject it. Line ranges over 256 MiB are rejected, Code Mode can start when telemetry or feature flags are disabled, and scripts can use their own `log`, `preview`, or `sleep` definitions.

    Usage details

    Who can use it: Claude Code 2.1.329; Code Mode and command/HTTP/MCP hooks.

    How to use it: To stop the operation, add `onFailure: "block"` to the relevant hook configuration; start Code Mode as usual.

    Tip: Split large file reads into chunks to reduce the risk of hitting the memory limit.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Code Mode can start even when OpenTelemetry is enabled

    Code Mode can now be used with OpenTelemetry events and metrics enabled. It sends events for tool calls made by Code Mode; tracing remains disabled.

    Usage details

    Who can use it: Claude Code 2.1.329; setups using OpenTelemetry events and metrics.

    How to use it: Leave your OpenTelemetry events and metrics settings enabled and start Code Mode using your existing activation method.

    Tip: Check event volume and your sensitive data policy before monitoring in production.

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

    Gateway visibly warns on every startup when `allow_cidrs` is empty

    The startup warning shown when `access_control.allow_cidrs` is empty is now repeated at every startup as a boxed notice in text logs. The notice explicitly warns against exposing the gateway to the public internet.

    Usage details

    Who can use it: Claude Code deployments using Gateway.

    How to use it: Check `access_control.allow_cidrs` in the Gateway configuration; define the permitted CIDR ranges and restart.

    Tip: Define your organization's CIDR ranges before exposing the Gateway to the internet in production.

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

    Permission, sandbox, and auto mode edge cases are handled more safely

    Auto mode no longer asks for unnecessary manual approval when deleting a symlink outside the workspace or one that points to itself; Linux `/sandbox` also clarifies that `allowUnixSockets` entries do not actually open sockets. False record errors caused by unreadable local settings files when adding permission rules have also been fixed.

    Usage details

    Who can use it: Claude Code 2.1.329; auto mode, Linux sandbox, and permission settings.

    How to use it: Inspect the active boundaries with `/sandbox`, and check the `/status` or `/config` message when adding a permission rule.

    Tip: Verify the target path before trusting auto mode with destructive operations such as deleting symlinks.

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

    macOS computer use applies stricter safety rules and a 30-second launch timeout

    macOS computer use has adopted new safety rules that reject certain actions and limit access to some apps. An app launch that does not complete is abandoned after 30 seconds.

    Usage details

    Who can use it: Claude Code 2.1.329; macOS computer use.

    How to use it: Sign in to a session using computer use on macOS; follow the safety message shown when an action is rejected or an app takes too long to launch.

    Tip: For apps that take a long time to launch, manually confirm the app is ready before retrying the task.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • Plugin / IntegrationClaude Code

    MCP authentication, reconnection, and file types are more robust

    Cleared MCP authentication no longer returns through an ongoing token refresh, and HTTP servers with slow reconnections are retried. PDF, image, DOCX, XLSX, and PPTX files sent as octet-stream by MCP or websites are no longer mistakenly saved as `.bin`.

    Usage details

    Who can use it: Claude Code 2.1.329; MCP servers and web tools.

    How to use it: Clear the server’s authentication in `/mcp` and reconnect; try sending a file as usual.

    Tip: Wait for the `/mcp` status message while the server reconnects instead of repeatedly starting the login flow.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code self-hosted runner

    Self-hosted runner handles LFS, SSH aliases, and remote MCP startup more reliably

    Git LFS repositories can start with pointer files and fetch content with `git lfs pull`. Startup instructions have been improved for remote repositories using an SSH host alias, GitHub SSH certificate, or GHE.com user; remote MCP and runner startup status are also reported more accurately after a timeout.

    Usage details

    Who can use it: Claude Code self-hosted runner and repositories using Anthropic-managed git.

    How to use it: Update the runner; in an LFS repository, run `git lfs pull` after the session starts and follow the suggested commands for SSH errors.

    Tip: In runner health checks, verify the checkout and MCP tool list in addition to the HTTP response.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • Pricing / PlanClaude Code

    Usage credits and spending screens show the actual status

    The prompt to sign in again after following the usage credits link for Pro and Max users has been fixed. `/usage` and `/model` distinguish between credits being disabled, depleted, or at the monthly limit; the spend limit menu also no longer shows a changed limit if the server rejects it.

    Usage details

    Who can use it: Claude Code 2.1.329; Pro and Max users, and models using usage credits.

    How to use it: Open `/usage-credits`, `/usage`, or `/model`; check the limit status and the server confirmation in the spend limit menu.

    Tip: After opening the credits link, don't change models without checking the actual limit.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code

    Previously scheduled tasks must be rescheduled in newly logged-in sessions

    Most tasks saved by previous versions must be rescheduled in a newly logged-in session. Resuming the same session with `--resume` keeps those tasks running; Exit and stop tasks stops them on the next resume.

    Usage details

    Who can use it: Claude Code 2.1.329; logged-in sessions using scheduled tasks.

    How to use it: Reschedule the old task in `/tasks` in the new session; use Exit and stop tasks to stop it.

    Tip: Record the schedules of critical tasks separately before upgrading.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • APIClaude Code self-hosted runner

    `--repository-folder` rejects repository conflicts on a capacity-1 runner

    A self-hosted runner running with `--capacity 1` can use most folders in the base folder with `--repository-folder`. Startup is rejected if another repository session conflicts in the same folder.

    Usage details

    Who can use it: Claude Code self-hosted runner; `--capacity 1`.

    How to use it: Start the orchestrator with `--repository-folder <folder>` and don't send a second concurrent repository session to the same folder.

    Tip: Use fixed, unique folder names for each repository in single-capacity runner pools.

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

    `/mcp` shows the remote server’s JSON-RPC error message

    A new Server message row has been added to the server page for remote HTTP MCP servers that reject the connection. The row displays the JSON-RPC error message returned by the server directly.

    Usage details

    Who can use it: Claude Code 2.1.329 and remote HTTP MCP servers.

    How to use it: Open `/mcp`, select the relevant remote server, and read the error in the Server message row.

    Tip: Fix the method or authentication issue indicated in this message instead of relying on the generic connection error.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Selecting a session with `/teleport` no longer overwrites the open conversation

    When a cloud session is selected through `/teleport`, a new session starts without replacing the open conversation. The previous conversation can be restored with `/resume`.

    Usage details

    Who can use it: Claude Code 2.1.329 and cloud sessions.

    How to use it: Type `/teleport` and select a cloud session; use `/resume` to restore the current conversation.

    Tip: To preserve important local context, use this new-session behavior instead of closing the conversation before teleporting.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • APIClaude Code

    Automatic limit retry for `max_tokens` errors removed

    The automatic retry behavior added in 2.1.326, which retries using the limit specified in a `max_tokens` error, has been removed. If a gateway model supports a lower output limit, Claude Code no longer retries automatically with that value.

    Usage details

    Who can use it: Claude Code 2.1.329; gateway or custom model endpoints.

    How to use it: Find out the actual output limit of your gateway model and, if needed, set `CLAUDE_CODE_MAX_OUTPUT_TOKENS` to that limit.

    Tip: For gateways, explicitly specify model limits lower than the Anthropic default in your deployment configuration.

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

    Auto mode runs `git status` less often in large repositories

    In large repositories, Auto mode runs `git status` less often to avoid delaying most requests. Changes that delete the working tree may be blocked until the working tree is checked.

    Usage details

    Who can use it: Claude Code 2.1.329 and large Git repositories using Auto mode.

    How to use it: Use Auto mode as usual; if a deletion is blocked, check the working tree status and approve again.

    Tip: Treat this behavior as a safety check to prevent loss of changes, not as a delay.

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

    Some teammate panels now start with `/bin/sh`

    Some teammate panels start through `/bin/sh` instead of Bash or the user's interactive shell. As a result, shell startup files may not be read by every teammate process.

    Usage details

    Who can use it: Claude Code 2.1.329 and teammate panels.

    How to use it: Explicitly set the environment variables required by the teammate in its working environment rather than relying on a startup file.

    Tip: For tasks that depend on PATH or aliases, verify the teammate shell with a small diagnostic command.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • SettingClaude Code

    The trust prompt shows where folder hooks are defined

    The dialog shown when folder trust is requested now also identifies the source of hook definitions. This lets you see which settings file will run before you approve.

    Usage details

    Who can use it: Claude Code 2.1.329 and local sessions using the trust dialog.

    How to use it: Open a project folder in Claude Code; review the hook definition line in the trust dialog.

    Tip: Before granting trust, compare the hook path with the settings file you expect in the repository.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code

    macOS login refresh checks Keychain write access

    Login refresh on macOS now checks write access using a Keychain item named `Claude Code-write-check`. This is intended to reduce repeated logouts when Keychain cannot be read or written.

    Usage details

    Who can use it: Claude Code 2.1.329; macOS.

    How to use it: No settings changes are needed; use login refresh as usual after updating Claude Code.

    Tip: If the issue persists, check Claude Code’s write permission and lock status in Keychain.

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

    Plugin and Claude Docs connector connections recover more reliably

    An uninstalled plugin's MCP server no longer remains connected in a headless session. A bug that caused Claude Docs connector tools to disappear when a plugin or MCP server at the same address failed—and incorrectly suggested `/login` as a fix—has been corrected.

    Usage details

    Who can use it: Claude Code 2.1.329; headless sessions, plugins, and Claude Docs connector.

    How to use it: After updating the plugin list, run `/reload_plugins` or start a new session; check for Claude Docs tools in `/mcp`.

    Tip: Troubleshoot conflicts by disabling plugins and connectors that use the same address one at a time.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code self-hosted runner

    Self-hosted runner keeps clones from multiple hosts in separate paths

    A second host copy with the same owner/repository name is now stored under `_repos/clones/<host>/`. Repositories with unusual characters in their address paths are also cloned into the new folder; rules and volumes tied to the old folder are not applied to the new clone.

    Usage details

    Who can use it: Claude Code self-hosted runner.

    How to use it: Update the runner and check the relevant repository clone paths under `_repos/clones/<host>/`.

    Tip: Move rule and volume definitions to the new clone path for each host.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • SettingClaude Code self-hosted runner

    Self-hosted runner reports ignored default permission mode at startup

    Claude Code 2.1.328 added a warning at self-hosted runner startup that sessions ignore the default permission mode in the host's `settings.json`. This makes the difference between the administrator setting and actual session behavior visible.

    Usage details

    Who can use it: Claude Code self-hosted runner.

    How to use it: Start the runner and read the permission mode warning in the startup logs; define the active mode in managed settings.

    Tip: Verify the settings the runner actually applies in a test session instead of relying on the host setting.

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

    Team sharing and passes screens now reflect server results accurately

    If the server rejects a guide upload, `/team-onboarding` no longer shows the guide as shared. After a server error, `/passes` also no longer shows an incorrect pass count or the usage credits option when no guest passes remain.

    Usage details

    Who can use it: Claude Code 2.1.329; Team and Enterprise onboarding and passes flows.

    How to use it: Reopen `/team-onboarding` or `/passes`; compare the result with the server's confirmation.

    Tip: If sharing fails, check the organization's upload permissions before resending the invitation.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Terminal setup and IDE diff saves preserve file integrity

    `/terminal-setup` no longer breaks VS Code settings around comment blocks or loses the Shift+Enter binding when a string is missing. Accepting an IDE diff also preserves the file's final newline and BOM.

    Usage details

    Who can use it: Claude Code 2.1.329; developers using VS Code and terminal setup.

    How to use it: Run `/terminal-setup` and save an IDE diff with Accept; check the settings and the end of the file.

    Tip: For encoding- or BOM-sensitive files, verify the file's final bytes in CI after Accept.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Visibility fixed in terminal panels and accessibility flows

    Panels, questions, and rows that were cut off in short windows have been fixed in Fullscreen, `/help`, history search, and the background task list. Resumed sessions also no longer incorrectly show truncated edits, questions, and plans as Rejected or Declined.

    Usage details

    Who can use it: Claude Code 2.1.329; terminal interface and accessibility users.

    How to use it: Open `/tasks`, `/help`, or `/resume` in a short terminal window and check that the lists are complete and focus works.

    Tip: If lines are cut off when you narrow the terminal, enlarge the window; don't assume an unseen prompt was confirmed.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    Long sessions, Windows startup, and cloud plugin loading are faster

    Memory usage in long sessions with many files and startup/first search on Windows have been improved. Cloud sessions retry plugins that could not be downloaded at startup; Desktop Remote Control becomes ready faster after a response.

    Usage details

    Who can use it: Claude Code 2.1.329; Windows, Claude Desktop, and cloud sessions.

    How to use it: Update Claude Code; restart the cloud session and wait for the plugin to be retried automatically at the next startup.

    Tip: After an initial plugin download issue, wait for the next automatic retry instead of repeatedly restarting.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code for VS Code

    Button and reason visibility fixed in long VS Code permission prompts

    In Claude Code 2.1.328, clipping of the No button, response field, and title in short panels or with long commands, URLs, and descriptions was fixed. Reasons no longer disappear or get cut off midway in blocked or failed tool calls.

    Usage details

    Who can use it: Claude Code for VS Code.

    How to use it: Open a permission prompt containing a long command or URL; if the reason is not visible, enlarge the panel and use the latest extension.

    Tip: Make sure you can see the entire reason instead of responding to permission prompts in a narrow panel.

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

    An explicit warning appears in the conversation if a managed hook fails to load

    When hooks in Managed settings fail to load, Claude Code reports this in the conversation, in Claude Desktop, and in other non-interactive hosts. Instead of failing silently, it displays an error that an administrator can address.

    Usage details

    Who can use it: Claude Code 2.1.329; Claude Desktop and non-interactive Claude Code hosts.

    How to use it: Start a session that uses managed settings; if loading fails, check the warning in the conversation stream.

    Tip: When you see the warning, check your organization’s managed settings deployment first rather than changing local settings.

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

    zsh command parsing, compaction, and terminal streaming are more robust

    In Claude Code 2.1.328, zsh words beginning with `=` no longer incorrectly trigger command path expansion. Automatic compaction no longer incorrectly reuses the previous system prompt; issues with long or malformed Markdown, rare screen freezes, paste handling, and lost Esc input were also fixed.

    Usage details

    Who can use it: Claude Code 2.1.328; zsh and long terminal sessions.

    How to use it: Update Claude Code and rerun commands such as `echo ====` in zsh; use compaction as usual in long sessions.

    Tip: If the terminal connection is slow, check that the pasted command actually appears in the prompt before running it.

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

    `/status` explains why managed marketplaces are not updated automatically

    A Marketplace auto-update line has been added to `/status`. It names managed marketplaces that do not receive automatic updates and explains why.

    Usage details

    Who can use it: Claude Code 2.1.329 and installations using managed marketplaces.

    How to use it: Run the `/status` command; inspect the Marketplace auto-update line.

    Tip: Check the stated reason for the update restriction with your administrator.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code

    The task counter shows all background jobs in one total

    The `/tasks` header and footer now count all shells, monitors, MCP calls, agents, workflows, and scheduled jobs in a single total. In Stop and SubagentStop hooks, Monitor scripts are identified as `monitor`.

    Usage details

    Who can use it: Claude Code 2.1.329.

    How to use it: Run multiple background jobs and check the total in the `/tasks` header; look for the `monitor` label in hook logs.

    Tip: Remember that the total is not just the number of agents; shell and MCP jobs are included too.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • AvailabilityClaude Code

    Remote Control titles and rename/diff views are fixed

    In `claude remote-control` sessions, `/rename` and UserPromptSubmit hook titles are now visible on claude.ai and mobile. Remote Control diffs no longer list renamed and edited files twice.

    Usage details

    Who can use it: Claude Code Remote Control, claude.ai, and mobile clients.

    How to use it: Use `/rename` in a Remote Control session and check the title on claude.ai or mobile; open the diff for a renamed file.

    Tip: Don't evaluate files containing both renames and edits solely by their total line count.

    ✓ Verified by official sourceSource: Claude Code v2.1.329 sürüm notları
  • CapabilityClaude Code for VS Code

    VS Code permission navigation and hidden-character display have changed

    In VS Code permission prompts, Down now stops at the last item instead of wrapping to Yes. Hidden non-printable characters and control sequences in command output are shown as symbols; this may affect rendering in some complex scripts.

    Usage details

    Who can use it: VS Code changes documented in Claude Code 2.1.329.

    How to use it: Use Down to navigate to the last option in a VS Code permission prompt; inspect control characters shown as symbols in output that looks corrupted.

    Tip: Account for this display behavior when evaluating terminal rendering in Persian, Khmer, Lao, Sinhala, or Malayalam text.

    ✓ Verified by official sourceSource: Claude Code changelog

OpenAI · ChatGPT

#2 GPT-6.2 Astra#6 GPT-6.2 Sol#9 GPT-6.1 Terra
Full history (339) →

New in this edition · 20

  • AvailabilityCodex

    GPT-6.2 Sol and Astra added to Codex catalogs; Sol is now the default

    Codex 0.171.0 adds GPT-6.2 Sol and GPT-6.2 Astra to its model catalogs and sets Sol as the catalog's default selection. This is a new detail about default catalog behavior in Codex, separate from the November 17–18 model launch.

    Usage details

    Who can use it: Codex 0.171.0 and later; catalog and model access may depend on account, workspace, and backend permissions.

    How to use it: Update Codex to 0.171.0, open the model catalog, and select GPT-6.2 Sol or GPT-6.2 Astra; Sol is the default on the selection screen.

    Tip: Try Sol for quick everyday tasks and Astra for more complex reasoning and coding tasks, then choose a model based on the task.

    ✓ Verified by official sourceModels: GPT-6.2 Sol, GPT-6.2 AstraSource: Codex 0.171.0 release notes
  • FeatureCodex CLI

    Codex TUI makes workspace the default and adds document, image, and subagent previews

    Codex 0.171.0 makes the TUI workspace the default. Alongside the workspace, users can view searchable document previews, image previews, and subagent transcripts next to the main conversation.

    Usage details

    Who can use it: Codex 0.171.0 and later; Codex CLI users working with the TUI workspace.

    How to use it: Update Codex and launch TUI; open documents or images from the workspace sidebar, or navigate to the relevant transcript from subagent activity.

    Tip: For long tasks, keep the relevant document or subagent transcript open in the sidebar so you don’t disrupt the main conversation context.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • Plugin / IntegrationCodex CLI

    Codex Enterprise MCP sign-in fixed for IdPs without refresh tokens

    Codex 0.171.0 fixes an issue with Enterprise MCP sign-in for identity providers that issue ID tokens but not refresh tokens. This allows MCP connections to complete with these types of enterprise IdP configurations.

    Usage details

    Who can use it: Codex 0.171.0 and later; Enterprise workspaces and ID-token-only identity providers.

    How to use it: Update Codex and restart Enterprise MCP sign-in using the normal OAuth flow; if the IdP issues only an ID token, the new version should not require an additional refresh token.

    Tip: Before an enterprise rollout, verify MCP sign-in with test users on IdPs that both issue and do not issue refresh tokens.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • APICodex CLI

    Codex exec can connect to existing app server sessions with --remote

    Codex 0.171.0 adds the `codex exec --remote <ADDR>` option to start or resume a non-interactive session on an existing app server. This lets automations reuse remote or already-running Codex sessions.

    Usage details

    Who can use it: Codex 0.171.0 and later; CLI environments with an accessible Codex app server address.

    How to use it: Find the address of the existing app server and run `codex exec --remote <ADDR> <task>`; provide the address and task values for your session.

    Tip: In CI jobs, manage the app server address and credentials through secure environment variables instead of hard-coding them on the command line.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • SettingCodex CLI

    Codex adds configurable undo/redo shortcuts to the non-Vim composer

    Codex 0.171.0 adds configurable undo and redo shortcuts for the non-Vim prompt composer. The feature also covers restoring edits and large pastes.

    Usage details

    Who can use it: Codex 0.171.0 and later; non-Vim prompt composer.

    How to use it: Update Codex, then use or configure undo/redo shortcuts in the non-Vim composer; restore content from within the composer when a large paste or edit is undone.

    Tip: Start with a small test snippet to try the shortcuts before sending a long prompt.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • SettingCodex CLI

    Codex TUI adds selection and authoring of custom MCP status lines

    When `tui_extensions_statusline` is enabled, Codex lets you select and author a custom MCP status line provider from within the TUI. This makes it easier to customize the status line with workspace, usage, or task information from MCP.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI installations with `tui_extensions_statusline` enabled.

    How to use it: Enable the feature in config with `tui_extensions_statusline`, open `/statusline` in the TUI, and select or configure the MCP status line provider.

    Tip: Start with a short status line that shows only the necessary fields; extensive MCP discovery can make the TUI less readable.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • CapabilityCodex CLI

    Codex TUI adds /rc and /remote remote control commands

    Codex TUI can now manage local remote control with the `/rc` and `/remote` commands and display device pairing codes.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI.

    How to use it: Type `/rc` or `/remote` in TUI; follow the remote control and device pairing steps shown by the commands.

    Tip: Share pairing codes only with a device you trust, and only during the short pairing window.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • SettingCodex CLI

    Codex preserves permission profile settings when the working directory changes

    Codex 0.171.0 preserves selected permission profile settings when changing the working directory. This prevents sandbox permissions from unexpectedly disappearing when switching projects or directories in the same session.

    Usage details

    Who can use it: Codex 0.171.0 and later; working-directory flows that use permission profiles.

    How to use it: Update Codex, select a permission profile, and change the working directory; check in the TUI permissions view that the profile is preserved in the new directory.

    Tip: In enterprise environments, restrict permission profiles with managed requirements and separately test working-directory changes.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • CapabilityCodex CLI

    Codex prevents late-completing events from canceling approvals for new turns

    Codex 0.171.0 prevents a late completion event from an earlier turn from canceling a pending approval for a newer turn. This makes approval flows more reliable for tasks that are routed in succession or run concurrently.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI tasks that require approval.

    How to use it: Update Codex and start two consecutive turns that require approval; approve the new turn only through that turn's prompt.

    Tip: When multiple turns are active, review approval requests with their thread and turn context.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • CapabilityCodex CLI

    Codex preserves drafts and queued input in their original threads when switching sessions

    Codex 0.171.0 preserves drafts, queued follow-up messages, and truncated input along with their original threads when switching or resuming sessions. This prevents unfinished composer content from being moved to the wrong conversation.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI session-switching and resumption flows.

    How to use it: Update Codex; leave a draft or queued follow-up message in one session and switch to another, then return to the previous session and continue the content in the same thread.

    Tip: For long tasks, it's now safer to leave a draft in place rather than delete it before switching sessions; still, don't leave sensitive information queued.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • CapabilityCodex CLI

    Codex TUI fixes scrolling, caret, and redraw issues

    Codex 0.171.0 fixes stray characters appearing while scrolling, composer caret jumps, and redraws using outdated terminal dimensions. This improves TUI behavior, especially with heavy output and terminal resizing.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI.

    How to use it: Update Codex to 0.171.0; use scrolling through long output and resize the window in TUI without requiring any additional settings.

    Tip: If the issue persists after resizing the terminal, check your Codex and terminal versions before restarting the session.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • CapabilityCodex CLI

    Codex preserves previous diff edits when steering an active turn

    Codex 0.171.0 preserves previous edits in the aggregate diff view when steering an active turn. This lets new steering build on the same workflow without accidentally losing previous changes.

    Usage details

    Who can use it: Codex 0.171.0 and later; TUI workflows that steer an active turn.

    How to use it: Update Codex, use steering during an active turn, and check the aggregate diff view to confirm that previous edits remain.

    Tip: Review the diff after steering and resolve any unexpected conflicts before committing.

    ✓ Verified by official sourceSource: Codex 0.171.0 release notes
  • AvailabilityHealth in ChatGPT

    Health expands to Android Health Connect, desktop, dots, and Sites

    Health is expanding to include Health Connect on Android, as well as the ChatGPT desktop app, dots, and Sites. New use cases include reviewing sleep and activity patterns using connected health data, proactive dot suggestions, and creating a workout-tracking site.

    Usage details

    Who can use it: Signed-in Free, Go, Plus, and Pro users in the US; Android Health Connect and gradual rollout. The release notes specify that dots access is available to Pro users.

    How to use it: Update the ChatGPT app; open Health in the sidebar, or the Explore menu if it isn’t visible. On Android, connect Health Connect; on desktop, create a Site using Health data.

    Tip: Use @Health in conversations where you want health data to inform the response, and connect only the data categories you need.

  • AvailabilityDots in ChatGPT

    Dots expands with a small Plus test and to all Business Standard users

    According to an announcement by OpenAI’s Codex and ChatGPT leads, dots is beginning tests with a small number of ChatGPT Plus users. The announcement also says dots will be rolled out to all ChatGPT Business Standard users, with the rollout starting on November 19.

    Usage details

    Who can use it: A limited number of ChatGPT Plus users and all ChatGPT Business Standard users; the announcement does not specify a region.

    How to use it: Open the dots area in ChatGPT desktop or web; if the rollout has reached your account, create a new dot or select an existing one. If the option isn’t visible on your Plus account, you may not yet be included in the test group.

    Tip: Before giving dots broad permissions, narrowly define its goal, connected apps, and actions that require approval.

  • ModelOpenAI Python SDK

    gpt-live-1.1 added to Python SDK Live model options

    OpenAI Python SDK v3.56.0 adds the `gpt-live-1.1` identifier to Live model options. Developers can select the new Live model in the SDK’s session configuration.

    Usage details

    Who can use it: OpenAI Python SDK v3.56.0 and later; developers with Live API access.

    How to use it: Run `pip install --upgrade openai==3.56.0` and select `model="gpt-live-1.1"` in the Live session configuration.

    Tip: Before deploying to production, compare audio quality, latency, and cost with your existing `gpt-live-1` workflow.

  • CapabilityOpenAI Python SDK

    Python SDK adds standby support for Live WebRTC sessions

    OpenAI Python SDK v3.57.0 supports pausing Live WebRTC sessions when idle and resuming them when speech is detected. The new configuration covers a speech wake policy, VAD threshold, idle timeout, paused/resumed events, and active/paused billing seconds.

    Usage details

    Who can use it: OpenAI Python SDK v3.57.0 and later; developers using Live WebRTC.

    How to use it: Update the SDK; when creating a session, set the `wake_policy: {"type": "speech", "vad_threshold": 0.5}` and `idle_timeout_s` fields. Monitor `session.paused` and `session.resumed` events over a sideband or permitted data channel.

    Tip: Use standby for short, predictable silences, and raise the `vad_threshold` value carefully in noisy environments.

  • Plugin / IntegrationChatGPT

    ChatGPT adds a new plugin for Accounting workflows

    The Accounting plugin supports workflows for reconciling transactions from records and templates, drafting invoice entries, and generating month-end reports. It is used through the plugin in the configured ChatGPT desktop app with computer use.

    Usage details

    Who can use it: ChatGPT desktop app for macOS and Windows; subject to plan and workspace eligibility. Admin approval is required in Enterprise workspaces.

    How to use it: Open the ChatGPT desktop app, set up computer use, open the Accounting plugin list, and select Install plugin if available. In Enterprise, first request Accounting approval from your workspace admin.

    Tip: Standardize record templates and reconciliation rules before generating month-end reports; review the output before importing it into your accounting system.

  • CapabilityCodex CLI

    Codex CLI documents a sandbox command that runs commands under a permission profile

    OpenAI released the `codex sandbox` command, which makes the Codex sandbox mechanism available for use with standalone tools. The command can run an arbitrary command by applying file system and network restrictions from a named permission profile in config.toml; MCP servers can also be wrapped this way.

    Usage details

    Who can use it: Environments supported by Codex CLI; profiles can be defined in a user or organization config.toml file.

    How to use it: Define a permission profile in `config.toml`, then run `codex sandbox --permission-profile dev -- <command>`. To restrict an MCP server, wrap its command with the same sandbox call.

    Tip: Start with a read-only or network-free profile; expand write and domain permissions only after testing.

    ✓ Verified by official sourceSource: OpenAI — Inside the Codex sandbox
  • Plugin / IntegrationOpenAI Python SDK

    Python SDK adds Live MCP server configuration

    OpenAI Python SDK v3.55.0 supports MCP server configuration in Live API sessions. This lets Live-based applications configure MCP servers through the SDK's typed API surface.

    Usage details

    Who can use it: OpenAI Python SDK v3.55.0 and later; developers using the Live API.

    How to use it: Update the SDK with `pip install --upgrade openai==3.55.0` and use the MCP server configuration fields in the v3.55.0 types when creating a Live session.

    Tip: Configure MCP servers with least privilege and connect only the tools that the Live session needs for that flow.

    ✓ Verified by official sourceSource: OpenAI Python SDK v3.55.0 release
  • RetiredOpenAI Python SDK

    Python SDK hides Managed Agents budget controls and events

    OpenAI Python SDK v3.55.0 no longer exposes the control and event surfaces related to Managed Agents session budgets. This may require adaptation for applications that bind budget controls and events directly to SDK types.

    Usage details

    Who can use it: OpenAI Python SDK v3.55.0 and later; Python applications using Managed Agents.

    How to use it: Update the SDK to v3.55.0 and review code that relies on Managed Agents budget controls or budget events, moving it to supported API surfaces.

    Tip: Before updating, run backward compatibility tests for code paths that use budget events.

    ✓ Verified by official sourceSource: OpenAI Python SDK v3.55.0 release

Google · Gemini

#3 Gemini 4 Neon#4 Gemini 4 Carbon#8 Gemini 4 Hydrogen#10 Gemini 4 Argon
Full history (244) →

New in this edition · 16

  • APIGemini API

    Tool Retrieval in Gemini API loads tool schemas on demand

    The Interactions API can search for and load the schemas it needs instead of sending the full schemas for large tool libraries with every request. Tool Retrieval Preview provides server-side or client-side search to reduce input tokens and latency.

    Usage details

    Who can use it: Interactions API Preview; Gemini 3.6 and later, and all Gemini 4 models, are supported. Remote MCP servers can also use deferred tool loading.

    How to use it: Add the `{"type":"tool_search"}` tool to the request and set `defer_loading: true` for tools to defer. Define client-side search with `execution: "client"` for your own tool registry.

    Tip: Load frequently used tools immediately and defer tools in the long tail; verify token savings with your own traffic measurements.

  • FeatureGemini app for Mac

    Computer use automates apps and the web in Gemini app for Mac

    Gemini app for Mac can now perform multistep tasks such as clicking, typing, and filling out forms in desktop apps and websites in a separate window. Users choose app permissions, watch a live preview, and approve sensitive steps.

    Usage details

    Who can use it: In English in the US; requires an Apple Silicon Mac running macOS Sequoia 15.0 or later, a personal Google account for users aged 18+, and a Google AI Plus, Pro, or Ultra subscription.

    How to use it: Download Gemini app for Mac, enter your Computer use request, choose the apps to use, and follow the steps in the live preview; approve sensitive steps such as signing in or making a payment.

    Tip: Allow access only to the apps you need for long tasks, and make your first attempt a reversible task.

  • AvailabilityGemini API

    Genie 4 opens to developers in Gemini API and Google AI Studio

    Genie 4 is now available through the Interactions API and Google AI Studio. It can generate interactive worlds at 24 fps and 720p from text or a reference image, with synchronized natural audio and live keyboard/text controls; the API release notes tie the feature to the paid tier.

    Usage details

    Who can use it: Gemini API’s Interactions API and Google AI Studio; paid tier for developers.

    How to use it: Open the Genie 4 workspace in AI Studio or use `models/genie-4` in an Interactions API request; provide text or a reference image and guide the world with live controls.

    Tip: Clearly define the world’s rules and changeable elements in the initial prompt; making live changes in small steps helps preserve scene consistency.

  • Plugin / IntegrationAgent Registry

    Agent Registry adds policy and HITL metadata for MCP tools

    Agent Registry supports protocol-level metadata in MCP tool registrations. This includes IAP policy hints, human-in-the-loop controls, and provider metadata attributes; the feature is in Preview.

    Usage details

    Who can use it: MCP servers registered in Google Cloud Agent Registry; Preview.

    How to use it: Register the MCP server as a Service resource in Agent Registry, then use policy, approval, and provider metadata in tool definitions that conform to the `toolspec.json` schema.

    Tip: Make HITL controls the default security layer for sensitive tools, and test IAP hints alongside actual IAM policies.

  • FeatureGemini app

    Preferences and chat history from other AI apps can be imported into Gemini

    Gemini supports importing a preference summary and chat history from another AI assistant. Users can transfer their previous context by uploading a ZIP archive from their former provider to Gemini settings.

    Usage details

    Who can use it: In the Gemini app; no country or plan restrictions are specified in the source.

    How to use it: Open the Gemini import settings page and copy Google’s prompt; generate a preference summary in your former AI tool and paste it into Gemini. Export a chat ZIP from your former provider and upload it in Gemini settings.

    Tip: First review only the preference summary and remove personal or confidential information, then upload the chat ZIP.

  • CapabilityBigQuery Data Engineering Agent

    Observability for BigQuery Data Engineering Agent reaches GA

    Data Engineering Agent now lets you monitor token consumption, request volume, and model latency in Cloud Monitoring, and execution traces and tool calls in Cloud Trace. The feature is generally available.

    Usage details

    Who can use it: BigQuery Data Engineering Agent; Cloud Monitoring and Cloud Trace integrations; GA.

    How to use it: Enable agent observability for Data Engineering Agent, then view metrics in Cloud Monitoring and execution/tool traces in Cloud Trace.

    Tip: Start by adding model latency and token consumption to dashboards; track cost increases alongside tool-call volume.

  • AvailabilityBigQuery

    BigQuery generative AI functions support Gemini 4 Neon

    BigQuery generative AI functions now support the `gemini-4-neon` model, which is generally available. Neon can now be selected for generative AI functions in BigQuery.

    Usage details

    Who can use it: BigQuery generative AI functions; Gemini 4 Neon; GA.

    How to use it: Set `gemini-4-neon` as the model in the supported model selection for your BigQuery generative AI function call, and validate the query with a small data sample.

    Tip: Use Neon for queries that require complex reasoning; compare total query costs for high-volume, simple tasks.

    ✓ Verified by official sourceModels: Gemini 4 NeonSource: Google Cloud sürüm notları – 19 Kasım 2026
  • FeatureBigQuery

    Natural language search with CREATE AI INDEX in BigQuery enters Preview

    BigQuery supports creating AI indexes for natural language search across table columns. Users can now use `AI.SEARCH` without manually generating embeddings or managing models; the feature is in Preview.

    Usage details

    Who can use it: BigQuery; `CREATE AI INDEX` and `AI.SEARCH`; Preview.

    How to use it: Run `CREATE AI INDEX` for the relevant columns in BigQuery, monitor index preparation, and run natural language queries with `AI.SEARCH`.

    Tip: Apply the index only to text columns you want to search, and measure search quality with a small test set of real user queries.

  • AvailabilityCodeMender

    CodeMender supports Gemini 4 Hydrogen in Preview

    CodeMender can now work with Gemini 4 Hydrogen. Gemini 3.8 Flash remains the default if no model is selected.

    Usage details

    Who can use it: CodeMender Preview; `cm find`, `cm verify`, and `cm fix` commands; Gemini 3.8 Flash is the default model.

    How to use it: Add `--model gemini-4-hydrogen` to the relevant command; for example, run `cm find --model gemini-4-hydrogen` or `cm fix --model gemini-4-hydrogen`.

    Tip: First, run a small comparative scan on the same repository with the default model and Hydrogen, and measure accuracy and runtime.

    ✓ Verified by official sourceModels: Gemini 4 HydrogenSource: Gemini Enterprise Agent Platform sürüm notları – 19 Kasım 2026
  • ModelDocument AI Custom Extractor

    Gemini 3.1 Flash Lite-based model for Document AI Custom Extractor reaches GA

    The `pretrained-foundation-model-v3.1-lite-2026-10-08` version for Document AI Custom Extractor is generally available with the Gemini 3.1 Flash Lite LLM. This processor version does not support fine-tuning.

    Usage details

    Who can use it: Document AI Custom Extractor; GA; ML processing support is noted in the US and EU; no fine-tuning.

    How to use it: Create a Custom Extractor, select `pretrained-foundation-model-v3.1-lite-2026-10-08` as the processor version, and evaluate field extraction with sample documents.

    Tip: Choose this version if you don’t need fine-tuning; compare quality on custom-labeled data with your current processor version.

  • OtherGemini CLI

    Gemini CLI 0.70.0-preview.0 improves Unicode, worktree, and MCP security

    v0.70.0-preview.0 preserves Unicode prompts in session previews, strips ANSI sequences, preserves the selected worktree during authentication restarts, and fixes output token limits in the model list. It skips unsafe command permissions on system MCP servers and tightens some default permission behaviors.

    Usage details

    Who can use it: Gemini CLI v0.70.0-preview.0 preview release.

    How to use it: Install the release tagged `v0.70.0-preview.0` from the GitHub Releases page and try it in a preview environment; test backward-compatible behavior before using it in CI or production workflows.

    Tip: Recheck MCP servers after updating; unsafe command permissions may now be skipped automatically.

  • OtherCodeMender

    CodeMender v0.23.0 returns the correct error code for failed diff scans

    When a `cm find --diff` scan job fails, CodeMender v0.23.0 returns a nonzero exit code that CI can detect instead of incorrectly returning 0.

    Usage details

    Who can use it: CodeMender v0.23.0; available for use, especially in CI/CD environments.

    How to use it: Upgrade CodeMender to v0.23.0 and check the exit code of `cm find --diff` in your CI step.

    Tip: Keep an exit-code check in CI to explicitly stop on failed scans; don’t rely on the behavior of older versions.

  • AgentPersonalized Service Health

    Incident-correlation agent skill published for Personalized Service Health

    Personalized Service Health offers the `servicehealth-incident-correlation` agent skill, which makes it easier to report whether a Google Cloud incident affects a project and to share published symptoms and workarounds. The skill gives agents reusable reference and command information.

    Usage details

    Who can use it: Personalized Service Health and Google’s skills repository; skill installation depends on the user’s environment.

    How to use it: Run `npx skills add google/skills --skill servicehealth-incident-correlation` in the terminal and verify that your agent can use the skill.

    Tip: Add the skill to your incident response workflow, but verify generated impact and resolution reports against official Service Health records.

  • AvailabilityGemini Enterprise Agent Platform

    Gemini 4 Neon reaches GA on Gemini Enterprise Agent Platform

    Gemini 4 Neon is generally available on Agent Platform with a combined input/output limit of 1,048,576 tokens, long-form reasoning with `continuation_token`, LOW–XHIGH thinking levels, and persistent implicit caching. The model is available through GenerateContent, the Interactions API, and the OpenAI-compatible Chat Completions API on the global endpoint and in the us and eu multi-regions.

    Usage details

    Who can use it: Gemini Enterprise Agent Platform; `global`, `us`, and `eu`; supports function calling, code execution, computer use, URL context, and Agentic video understanding.

    How to use it: Select the `gemini-4-neon` model in requests; use `thinking_level="XHIGH"` for complex tasks, `continuation_token` for long outputs, and Interactions API `background=true` for background execution.

    Tip: Keep the `HIGH` default for most tasks and reserve `XHIGH` for difficult reasoning tasks; monitor implicit cache usage for long, repetitive prompts.

  • AvailabilityGemini Enterprise

    Gemini 4 Neon reaches GA in Gemini Enterprise apps and tools

    Gemini 4 Neon is generally available in Gemini Enterprise web and mobile apps, Workflow Builder, and AI developer tools. The model is off by default in the app; in Workflow Builder, it can be used for agent flow nodes and no-code agents.

    Usage details

    Who can use it: `global`, `us`, and `eu` multi-regions; an administrator must enable it in the web app, and it is available to licensed users in AI developer tools with the default Antigravity model selection.

    How to use it: Enable Gemini 4 Neon in the web app model availability settings in the Gemini Enterprise admin console. If you use a custom Antigravity model list, also select the model in AI developer tools settings; assign it to an agent node in Workflow Builder.

    Tip: First enable it for a small group of users in the organization and test `us`/`eu` data location requirements alongside endpoint selection.

  • APIAPI Gateway

    API Gateway model routers can route to different Google API hosts

    Model routers can now use Google API backends on different hosts and in different projects in the same routing configuration, such as the global `aiplatform.googleapis.com` and the regional `us-central1-aiplatform.googleapis.com`. All backends in new configurations must use HTTPS.

    Usage details

    Who can use it: API Gateway model routers; global/regional Google API endpoints and different projects.

    How to use it: In the model routing configuration, define backend hosts as separate Google API endpoints, use `https` for each backend, and verify permissions across projects.

    Tip: Choose regional endpoints based on latency or data location needs; test the new configuration in a separate gateway first rather than assuming existing deployments are unaffected.

Meta · Meta AI

Muse Wonder FullMuse WonderMuse Ember
Full history (128) →

New in this edition · 3

  • APIWhatsApp Cloud API

    WhatsApp Cloud API adds a quality rating and webhook event for business calls

    The Phone Numbers API now returns a directional calling quality rating for phone numbers used to make business-initiated WhatsApp calls. Values can be GREEN, YELLOW, or RED; rating changes are reported through CALLING_QUALITY_UPDATE events in the account_update webhook.

    Usage details

    Who can use it: Limited availability; according to the documentation, this applies to eligible business numbers with Colombian and Mexican country codes in the quality-based model. Existing eligible numbers with Calling API enabled are migrated automatically.

    How to use it: Enable Calling API for the business phone number; read the calling_quality_rating field in the Phone Numbers API response and handle the CALLING_QUALITY_UPDATE event in the account_update webhook.

    Tip: Don’t use GREEN/YELLOW/RED values as the sole decision criterion; track rating transitions over time and review call volume and user permissions after a YELLOW or RED rating.

  • AvailabilityForum

    Forum app launches on Android in Indonesia; iOS coming soon

    The dedicated Forum app for Facebook Groups is now available to download on Android in Indonesia; Meta says the iOS version is coming soon. The app brings groups together in one place and offers Ask, Your Algorithm, Top voice, and admin AI tools for group administrators.

    Usage details

    Who can use it: Indonesia; available now on Android through Google Play, with the iOS version coming soon. Users sign in with their Facebook accounts.

    How to use it: On an Android device in Indonesia, download the Forum app from Google Play, sign in with your Facebook account, and open your existing groups in the Forum feed; if you’re an administrator, try the admin AI tools within your group.

    Tip: Use Ask not just for general searches, but also to compare real user experiences in the groups you’ve joined; always review admin AI’s approval and draft suggestions before publishing them.

  • APIWhatsApp Cloud API

    `free_tier` added to WhatsApp status webhooks for the free service allowance

    The `free_tier` value has been added to the subtype field of the Pricing object. Service messages delivered within a business phone number’s monthly free service message allowance carry this value, while `paid_exempt`, used for eligible government and nonprofit organizations, remains unchanged.

    Usage details

    Who can use it: For businesses using WhatsApp Cloud API status messages webhooks; appears on deliveries covered by the monthly free service message allowance.

    How to use it: Update your messages status webhook parser to accept the free_tier value in the pricing.subtype field; report free_tier and paid_exempt separately, and evaluate them alongside the billable field in your billing logic.

    Tip: Instead of rejecting new subtype values with a strict enum, log unknown values and handle them in a backward-compatible way; this prevents future pricing signals from disrupting the webhook flow.

Moonshot AI · Kimi

Kimi K3.1 Pro
Full history (320) →

New in this edition · 40

  • Plugin / IntegrationKimi Code Desktop

    Desktop can connect to Claude Code, Codex, and Cursor sessions

    Kimi Code Desktop 1.4.0 adds the ability to connect to external Agents such as Claude Code, Codex, and Cursor, hand off sessions, review history, and collaborate across turns.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0; developers using Claude Code, Codex, or Cursor.

    How to use it: Update to Desktop 1.4.0 and select the Claude Code, Codex, or Cursor session you want to connect to from the external Agent connection/session handoff interface.

    Tip: Before handing off, check that the project root and active session are correct; shared history may contain sensitive command output.

  • CapabilityKimi Code Desktop

    Background task notifications are now collapsible rows

    Background task notifications in Desktop are now displayed as collapsible rows, like tool calls; task duration and completed/failed/stopped status remain clearly visible.

    Usage details

    Who can use it: Kimi Code Desktop users.

    How to use it: Run a background task, then expand its row in the conversation stream to review the details, duration, and final status.

    Tip: For long-running tasks, open only failed or long-running rows to keep the conversation stream more readable.

  • AvailabilityKimi Code Desktop

    Desktop messages sent during a running turn now steer by default

    In Desktop 1.4.0, messages sent while Agent is running now steer the current turn immediately by default, with a tip shown on first use. Users can switch back to the previous queueing behavior in Settings; this extends to Desktop a behavior setting that was already available in TUI.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0 users; the existing TUI setting is already recorded in the known list, and this reports its availability on a new platform.

    How to use it: Send a message while Agent is running and see it steer the current turn. To queue messages, set the follow-up message behavior during a running turn to `Queue` in Settings.

    Tip: Using steer for quick corrections and Queue for separate follow-up tasks gives more predictable results.

  • CapabilityKimi Code CLI

    Compaction output for Claude Opus 5 and Haiku 5 capped at 128,000 tokens

    For Claude Opus 5.5 and Haiku 5.5 used through the Anthropic-compatible endpoint, automatic compaction requests can no longer exceed the model’s 128,000-token single-response limit. Smaller explicit `max_output_size` values are preserved.

    Usage details

    Who can use it: Users of these Claude model families with an Anthropic-compatible provider in Kimi Code CLI/Desktop, ACP, and SDK.

    How to use it: Use `claude-opus-5-5` or `claude-haiku-5-5` with a custom Anthropic provider, then run the normal automatic compaction flow. If you set a lower `max_output_size`, that value is preserved.

    Tip: In sessions using a one-million-token context, planning earlier, controlled compaction can reduce interruptions caused by hitting the limit.

    ✓ Verified by official sourceModels: Claude Opus 5.5, Claude Haiku 5.5Source: Claude çıktı üst sınırı pull requestiSource: Kimi Code 2.7.0 CLI sürüm notları
  • FeatureKimi Code Desktop

    Rows, columns, or regions in Markdown tables can be selected and copied

    In the Desktop transcript, you can select cell regions, rows, or columns in Markdown tables. Copy a selection by right-clicking or using a keyboard shortcut.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0 transcript users.

    How to use it: Open a Markdown table in the Transcript; select a cell region, row, or column, then right-click or use the supported copy shortcut.

    Tip: Selecting only the required row or column from model output is faster than cleaning up the entire table.

  • FeatureKimi Code Desktop

    Sessions can be marked as unread, and projects can be pinned to the sidebar

    Added `mark as unread/read` actions to the session context menu. Projects can now appear alongside sessions in the pinned section and be reordered by dragging.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0 users.

    How to use it: Right-click a session in the sidebar and select `Mark as unread/read`; use the pin action on a project, then drag it to reorder it in the pinned section.

    Tip: Pinning projects that need follow-up and marking completed sessions as read makes long project lists quicker to scan.

  • Plugin / IntegrationKimi Code Desktop

    Qwen Coding Plan added to global and China providers

    Added both global and China site options for Qwen Coding Plan to the Desktop provider selection. Qwen subscriptions can now be configured alongside other selectable model providers.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0; global or China-region users of Qwen Coding Plan.

    How to use it: Open the provider/model selector in Desktop, choose the `Qwen Coding Plan` provider that matches your region, and complete authentication.

    Tip: Choose the global or China provider that matches your account’s region; using the wrong endpoint may cause authentication to fail.

  • CapabilityKimi Code Desktop

    Local projects and older sessions remain visible even if the SSH Host doesn’t respond

    When the SSH Host is down, the local project list no longer becomes empty or gets stuck waiting for a timeout. Loading older project sessions in the sidebar has also been fixed; if remote discovery takes more than five seconds, the available local/reachable results are shown.

    Usage details

    Who can use it: Users of Kimi Code Desktop/Web who use SSH Host.

    How to use it: Update to Desktop 1.4.0 and open the project sidebar as usual. If the Host is unreachable, use the local results and reload when the Host is back.

    Tip: On networks where the remote Host is slow, it’s faster to use local projects for the initial view and refresh remote projects later.

  • FeatureKimi Code CLI

    The TUI `/feedback` window now accepts multiline text

    `/feedback` input now preserves line breaks and supports multiline editing. `Shift-Enter` or `Ctrl-J` adds a new line, while `Enter` submits; the attached-file selection and upload flow are unchanged.

    Usage details

    Who can use it: Users submitting feedback in the Kimi Code CLI TUI.

    How to use it: Type `/feedback`; enter or paste multiline text, use `Shift-Enter`/`Ctrl-J` for a line break, `Enter` to submit, and `Esc` to cancel.

    Tip: Write the issue, reproduction steps, and expected result on separate lines to help the support team diagnose the problem.

  • Plugin / IntegrationKimi Code CLI

    Vertex AI can now use Google ADC without an API key

    If no explicit API key is set, the `vertexai` provider now accepts authentication through Google Application Default Credentials. The `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION` values in the provider environment are now passed to the Google SDK.

    Usage details

    Who can use it: print, KAP, ACP, and CLI workflows using Vertex AI; users signed in with `gcloud` ADC.

    How to use it: In `config.toml`, add `GOOGLE_CLOUD_PROJECT` and `GOOGLE_CLOUD_LOCATION` under `[providers.<ad>.env]` for the `type = "vertexai"` provider; run `gcloud auth application-default login` once.

    Tip: If the project/region values in the shell conflict with TOML, use the TOML configuration for the request target.

  • CapabilityKimi Code Desktop

    Unread/read status no longer duplicates across consecutive steer messages

    When multiple steer messages are sent in a row, the unread/read status of earlier messages no longer multiplies or duplicates incorrectly. Message status and the progress indicator remain more consistent.

    Usage details

    Who can use it: Users who send consecutive messages to a running Agent in Kimi Code Desktop 1.4.0.

    How to use it: Send multiple steer messages while the Agent is running, then check that each message has a single read status in the transcript.

    Tip: Send small, prioritized corrections in sequence instead of splitting the same turn with too many steer messages.

  • CapabilityKimi Code Desktop

    Markdown table copy/expand buttons are no longer clipped in narrow windows

    When Side Chat is open and the transcript area gets narrower, Markdown table actions now move within the visible bounds. The `Copy table` and expand-table buttons remain clickable.

    Usage details

    Who can use it: Kimi Code Desktop/Web users with Side Chat open or using a narrow window.

    How to use it: Open Side Chat and hover over a Markdown table; use the copy and expand buttons directly in the visible area.

    Tip: In narrow layouts, copy the necessary portion before expanding the table; the buttons remain visible even when scrolling horizontally.

  • CapabilityKimi Code Desktop

    Model/permission popups no longer overflow in narrow and mobile layouts

    Fixed an issue that caused model and permission popups to extend beyond the screen in narrow or mobile-sized layouts. Selection and confirmation controls remain visible.

    Usage details

    Who can use it: Users of Kimi Code Desktop 1.4.0 in a narrow window or mobile-sized Web view.

    How to use it: Update the app and open the model or permission selector in a narrow window; select options directly from the visible window.

    Tip: If you can’t see the options on a narrow screen, update to 1.4.0 before changing browser zoom.

  • CapabilityKimi Code Desktop

    The Desktop session status dialog shows the current session ID

    The session status dialog now clearly displays the current session ID. This makes it easier to copy the ID for support, KAP automation, and opening the same session in another client.

    Usage details

    Who can use it: Kimi Code Desktop 1.4.0 users.

    How to use it: Open the session status dialog and copy the displayed session ID to use in a support ticket or KAP request.

    Tip: Adding the session ID to a technical support ticket instead of a user message helps support find the right session faster.

  • CapabilityKimi Code Desktop

    Screenshot magnifier and color picker take up less space

    The magnifier and color picker in Desktop screenshot tools have been made smaller, so they cover less of the image.

    Usage details

    Who can use it: Users of screenshot tools in Kimi Code Desktop 1.4.0.

    How to use it: In the screenshot flow, open the magnifier or color picker and use the new, smaller floating tools directly over the image.

    Tip: With the smaller tools, more of the area around the target pixel remains visible, making it easier to verify color selection.

  • CapabilityKimi Code Desktop

    Image upload failures no longer trigger unnecessary authentication rate limits

    An issue that caused the app to mistakenly count a failed image upload as an authentication attempt and trigger an auth rate limit has been fixed. This reduces false sign-in/limit errors in sessions that use images.

    Usage details

    Who can use it: Users adding images in Kimi Code Desktop 1.4.0.

    How to use it: Update Desktop and add the image again; sign in again if there is a genuine authentication error, but avoid unnecessary retries when only the image upload has failed.

    Tip: If the image upload issue persists, check the file size and network access; don’t repeatedly sign out and back in, which can increase rate limiting.

  • Plugin / IntegrationKimi Code Desktop

    An update button is shown for the partially installed WebBridge extension

    When the WebBridge extension is in `partial` state and a newer version is available in the catalog, the extensions page no longer hides the update button. The button runs the same safe and idempotent install/update flow.

    Usage details

    Who can use it: Users whose WebBridge installation in Kimi Code Desktop is incomplete and for whom a newer WebBridge version is available in the catalog.

    How to use it: Open the extensions page in Desktop; if an update notice appears on the `WebBridge` row, click the `Update` button.

    Tip: After a partial installation, try this update flow before removing and reinstalling manually.

  • CapabilityKimi Code Desktop

    macOS IME candidate window no longer hides behind the screenshot overlay

    Fixed an issue that caused the IME candidate window to be hidden in the background while the screenshot overlay was open on macOS. Candidate selection remains visible during the screenshot flow when using CJK input.

    Usage details

    Who can use it: macOS users using Kimi Code Desktop 1.4.0 and a CJK input method.

    How to use it: Update Desktop, enable the CJK IME, and enter text with the screenshot overlay open; select candidates as usual.

    Tip: If the candidate window is not visible when typing in Chinese/Japanese/Korean, first close the screenshot overlay and check for an app update.

  • CapabilityKimi Code Desktop

    Already-closed sessions no longer wake up when stopping a session

    A bug that could cause other closed sessions to start running again when a session was stopped has been fixed. The stop action now applies only to the target session.

    Usage details

    Who can use it: Users managing multiple sessions in Kimi Code Desktop 1.4.0.

    How to use it: Use the stop action on a running session and check that the status of other closed sessions hasn’t changed.

    Tip: In multi-session setups, check the sidebar statuses after stopping a session to confirm that the wrong session hasn’t woken up.

  • FeatureKimi Code Desktop

    New folders can be created directly from the project folder picker

    The folder row and action buttons have been restored when selecting a folder for a new project or project directory; users can create a new folder from within the picker. The full path is visible on hover.

    Usage details

    Who can use it: Users creating projects in Kimi Code Desktop 1.4.0.

    How to use it: Open the Add new project or Select directory window; use the new folder action in the folder row and, if needed, hover to check the full path.

    Tip: Creating the workspace from the project picker instead of first creating it in a file manager reduces the chance of choosing the wrong parent folder.

  • CapabilityKimi Code Desktop

    Repeated cryptic error notifications for expired provider credentials reduced

    Repeated, unclear `action failed` notifications when opening the model selector with expired provider credentials have been fixed, making the error flow clearer.

    Usage details

    Who can use it: Users with expired provider credentials in Kimi Code Desktop 1.4.0.

    How to use it: Open the model selector; if the credentials have expired, renew the provider account or sign in again with `/login`.

    Tip: Checking provider credential status first, rather than repeatedly retrying the same action, prevents unnecessary error notifications.

  • SettingKimi Code Desktop

    Project and session titles and dynamic favicons can be configured in browser tabs

    Desktop’s built-in browser now displays dynamic favicons and can use project and current session information in tab titles. This view can be configured under `Settings > Appearance`; the preview is shown as a single tab.

    Usage details

    Who can use it: Users of the Kimi Code Desktop 1.4.0 built-in browser.

    How to use it: Turn the browser tab title preview on or off under `Settings > Appearance`; check the change in built-in browser tabs.

    Tip: When working across multiple projects, keeping project and session information visible reduces the risk of acting in the wrong context between tabs.

  • CapabilityKimi Code Desktop

    Browser annotations fixed for local HTML files

    Fixed an issue that prevented annotations from working when local HTML files were opened in the built-in browser. You can now review local output and share text or visual context with Agent more reliably.

    Usage details

    Who can use it: Users who open local HTML files in the built-in browser in Kimi Code Desktop 1.4.0.

    How to use it: Open a local HTML file in the built-in browser from the Transcript or file tree; select the annotation tool and mark the relevant area.

    Tip: Before sharing a local prototype, annotate the HTML directly to show Agent which area needs changes.

  • CapabilityKimi Code Desktop

    The built-in browser no longer incorrectly shows an unsafe warning on some sign-ins

    An issue that caused the built-in browser to be marked as unsafe when signing in to some sites has been fixed. This resolves false warnings and sign-in flow issues without disabling a real security check.

    Usage details

    Who can use it: Users signing in to sites with the built-in browser in Kimi Code Desktop 1.4.0.

    How to use it: Update Desktop to 1.4.0 and try signing in to the site again in the built-in browser; continue to heed genuine certificate or site security warnings.

    Tip: Don’t use the fixed browser warning as a reason to ignore all security warnings.

  • AgentKimi Code CLI

    `/tasks` can also open foreground Agent activity

    In TUI, `/tasks` now lists running foreground subagents, and `Enter` opens their body, tool activity, and visible reasoning details. The main Agent continues working; once the task finishes, the open details continue to show the final state.

    Usage details

    Who can use it: Kimi Code CLI TUI users; in normal and full-screen TUI.

    How to use it: Type `/tasks` in TUI, select a `running` foreground Agent row, and press `Enter` to open the details.

    Tip: For a long-running subagent task, use this panel to track tool-call progress without interrupting the main conversation.

  • APIKimi Code KAP Server API

    WebSocket-free HTTP control flow for KAP released

    `kimi web --json` returns process information as JSON and generates an in-memory bearer token. The addition of the `GET /api/v1/sessions/{id}/interactions`, `POST /api/v1/sessions/{id}:wait`, and `GET /api/v1/kap.md` endpoints lets another Agent monitor Kimi Code, view approvals, and wait using HTTP only.

    Usage details

    Who can use it: Developers building Agent-to-Agent automation with KAP v1; `timeout_ms` range for `:wait` is 0–60000 ms.

    How to use it: Get the `origin`, `ws_url`, and `token` values from the first JSON line of the `kimi web --json` output; use the bearer token to call `GET .../interactions`, `POST ...:wait -d '{"timeout_ms":15000}'`, and `GET /api/v1/kap.md`.

    Tip: In a pure HTTP controller, call `:wait` first, then read pending interactions; also use the turns endpoint to get the active flow in real time.

  • RetiredKimi Code KAP Server API

    KAP REST API v2 session and MCP routes consolidated under v1

    The `/api/v2/sessions*` and `/api/v2/mcp/*` routes have been removed; session lists, grouping, Git information, bulk archiving, and MCP management are now consolidated under `/api/v1`. The old v1 live MCP view and `:restart` action have also been removed, so legacy integrations may receive 404s.

    Usage details

    Who can use it: Kimi Code Desktop/Web and KAP REST clients; legacy v2 integrations must migrate to the v1 contract.

    How to use it: Use `/api/v1/sessions` for session lists; switch to `/api/v1/sessions:archive` and `:restore` for archiving/restoring, and `/api/v1/mcp/*` routes for MCP management. Use `archived` instead of `include_archive`, and `has_prompt` instead of `exclude_empty`.

    Tip: As there is no backward-compatibility layer, migrate your client from v2 to v1 alongside version detection.

  • APIKimi Code KAP Server API

    KAP v1 session protocol unifies interrupt, profile, and flat event contracts

    KAP v1 standardizes a single session view, an endpoint for interrupting a specific turn, the side Agent lifecycle, flat `session`/`workspace`/`config`/`invalidated` events, and target editing through profiles. Legacy fragmented event envelopes and some older session routes have been removed; clients should migrate to the new fields.

    Usage details

    Who can use it: Kimi Code Web/Desktop and KAP v1 REST/WebSocket clients; the experimental protocol may change between versions.

    How to use it: Update your client to read `SessionInfo.activity`, `last_turn_outcome`, and the top-level `cwd` fields according to the updated KAP reference; use the `POST /api/v1/sessions/{session_id}/agents/{agent_id}/turns/{turn_id}:interrupt` contract to interrupt a turn.

    Tip: For live integrations, treat the running version’s `GET /openapi.json` and `GET /asyncapi.json` schemas as the source of truth rather than the documented page.

  • Plugin / IntegrationKimi Code CLI

    Amazon Bedrock provider and AWS credential chain support added to Kimi Code CLI

    Kimi Code can now connect to Amazon Bedrock models using the `type = "bedrock"` provider. Instead of a long-lived Bedrock key, a short-lived bearer token can be generated using AWS’s standard credential chain; the model protocol can be selected automatically or specified explicitly.

    Usage details

    Who can use it: Kimi Code CLI 2.7.0-canary and 2.7.0; users with an AWS account and region configured.

    How to use it: Define `[providers.bedrock] type = "bedrock"` in `config.toml`; set `AWS_REGION` or `AWS_DEFAULT_REGION` and sign in using AWS SSO, a shared profile, or an instance role. If needed, add `protocol = "anthropic"`, `"openai"`, or `"openai_responses"` to the model alias.

    Tip: A missing region definition stops requests explicitly rather than sending them to the wrong global endpoint; pin the region in the provider environment in production.

  • AvailabilityKimi Code CLI

    Full Kimi Code server and independent sessions enabled on SSH Hosts

    Selecting an SSH Host lets you run the full Kimi Code server, project, and Agent on a remote machine. Each Host uses its own Kimi account; the remote session continues running even if the local computer goes to sleep or the connection drops, and users can optionally copy custom model providers once.

    Usage details

    Who can use it: Kimi Code CLI/Desktop/Web and users configuring SSH Hosts; the remote Host must have its own Kimi login.

    How to use it: Configure the Host and select the SSH machine when creating a project; sign in to Kimi separately on the Host. Use `kimi server` for a headless local server, or `kimi server --open` to open the browser too.

    Tip: Instead of copying API keys automatically, transfer only the custom providers you need, with explicit approval.

  • CapabilityKimi Code ACP

    Local and SSH tool contexts are preserved when the ACP preview closes

    The ACP `session/load` and `session/close` flow no longer permanently replaces the actual local/SSH runtime environment with an `acp:*` environment. After the preview closes, Bash, Edit, Write, and similar tools continue to run in the same environment; SSH sessions also no longer accidentally fall back to the local machine.

    Usage details

    Who can use it: Users who open past local or SSH Kimi Code sessions with ACP clients.

    How to use it: In your ACP client, open the existing session with `session/load` and send `session/close` when you’re done; no additional environment restoration or manual `wire.jsonl` editing is required.

    Tip: You can also close ACP in SSH sessions while disconnected; the actual environment preserves the reconnection path for next time.

  • CapabilityKimi Code Bash tool

    Bash subprocesses receive session and workspace IDs

    Bash tool subprocesses now receive the read-only environment variables `KIMI_CODE_SESSION_ID` and `KIMI_CODE_WORKSPACE_ID`. Scripts, hooks, and Skills can therefore make API calls for the correct session/workspace without relying only on the working directory.

    Usage details

    Who can use it: Bash calls in TUI, print, Desktop, Web, ACP, and SDK.

    How to use it: Read `$KIMI_CODE_SESSION_ID` and `$KIMI_CODE_WORKSPACE_ID` in your Bash script and use them to associate KAP session creation or association requests with the correct IDs.

    Tip: If you use an explicitly managed workspace, include the workspace ID rather than creating a new session using only `$PWD`.

  • CapabilityKimi Code CLI

    Sessions open in another client can now be resumed in read-only mode

    `resume` can now load history while another Kimi Code client holds the session’s write lock. Sending messages and other write operations still require the lock to be released or a takeover; the change only enables reading history.

    Usage details

    Who can use it: Kimi Code CLI, SDK, and users who want to view a session open in another client.

    How to use it: Run `kimi --resume <session-id>` or open the session from the Sessions selector; if the write lock is still held, wait for the other client to release the session before sending a new message.

    Tip: Open the same session in two clients for observation only; do not expect an automatic takeover for concurrent writes.

  • APIKimi Code KAP Server API

    Endpoint added to close BTW side Agents in the KAP API

    KAP clients can now stop and remove only a `btw` side Agent from the record using `POST /api/v1/sessions/{session_id}/agents/{agent_id}:close`. The server clears pending interactions, publishes an `agent.status = "ended"` event, and returns `{ closed: true }`.

    Usage details

    Who can use it: Web/Desktop and external Agent clients using KAP v1; the main Agent and regular subagents cannot be closed with this endpoint.

    How to use it: Save the `agent_id` from the BTW creation response and, when closing, send `POST .../agents/{agent_id}:close` to the relevant session path; wait for the `ended` event over WebSocket.

    Tip: Calling this endpoint when closing the panel prevents invisible side Agents and pending approvals from piling up.

  • APIKimi Code KAP Server API

    API added to resume interrupted or failed turns without a fake message

    The new `POST /api/v1/sessions/{session_id}:continue` action resumes an incomplete, canceled, or failed main Agent turn without adding a new user message. Running sessions return `40901`; turns that cannot be resumed return `40940`.

    Usage details

    Who can use it: KAP v1 REST clients and automations; the new endpoint is available in Kimi Code 2.7.0 canary.

    How to use it: After observing that the session has stopped, send `POST /api/v1/sessions/{session_id}:continue`; if the response is `{ continued: true }`, track the result through the event stream.

    Tip: Use this action instead of a new user prompt so the model does not repeat steps it has already completed.

  • Plugin / IntegrationKimi Code Remote Control

    Remote Control WebSocket connections now use corporate proxy variables

    The Remote Control relay WebSocket now respects the `HTTP_PROXY`, `HTTPS_PROXY`, `ALL_PROXY`, and `NO_PROXY` variables. An invalid proxy URL prints a warning to stderr and falls back to a direct connection.

    Usage details

    Who can use it: CLI/Desktop users running Kimi Code Remote Control behind a corporate proxy.

    How to use it: Before starting Remote Control, set the appropriate `HTTPS_PROXY` or `HTTP_PROXY` value, and set `NO_PROXY` for relay domains that should be accessed directly; then run `kimi rc` or `kimi remote-control`.

    Tip: Verify that the proxy allows CONNECT/WebSocket traffic and that the `NO_PROXY` list does not inadvertently include the relay domain.

  • CapabilityKimi Code KAP Server API

    Grouping, ID projection, and sorting fixed in SSH Host session lists

    For queries that include multiple SSH Hosts, `view=by_workspace` now returns `groups` even without pagination parameters. The `fields=id,archived` projection and requested sort orders such as `updated_at_asc` and `created_at_desc` also no longer produce a 50001 error or silently fall back to the default sort order.

    Usage details

    Who can use it: Kimi Code Desktop/Web and KAP API clients using `hosts=include`.

    How to use it: Use `GET /api/v1/sessions?hosts=include&view=by_workspace&group_page_size=5` for a grouped list; for a lightweight list, use `fields=id,archived` and your preferred `sort` parameter.

    Tip: For large Host clusters, control response size by using the grouped view and lightweight ID projection in separate requests.

  • CapabilityKimi Code CLI

    `/btw` automatically reloads the session after an update from another client

    After another client updates the same session, `/btw` now reloads the session and retries once instead of stopping with a `session.stale` error. Other types of conflicts are reported as before.

    Usage details

    Who can use it: TUI and SDK users who use a session with multiple Kimi Code clients.

    How to use it: Run `/btw <question>` in TUI; if an external update is detected, wait for the response after the automatic reload completes.

    Tip: If you use parallel clients in the same session, try BTW’s automatic recovery before using manual `/reload`.

  • APIKimi Code KAP Server API

    Host product name added to server metadata and instance record

    The Kimi Code server now adds the Host product name it is running to the `~/.kimi-code/server/instances/*.json` record and the `GET /api/v1/meta` response. This lets clients distinguish between multiple local/remote server types.

    Usage details

    Who can use it: Desktop/Web/automation clients using the KAP API and server instance records.

    How to use it: Call `GET /api/v1/meta` with the server token and use the returned product field alongside the instance records; support older server versions if the field is missing.

    Tip: In multi-Host interfaces, use this product information instead of the displayed name to distinguish server-type capabilities.

  • AvailabilityKimi Code CLI

    24-hour cap removed from phased CLI update plans

    The client now applies phased update delays published in `latest.json` that exceed 24 hours without capping them. Users outside the rollout group are held back for at least 24 hours or the longest configured delay.

    Usage details

    Who can use it: Users on Kimi Code’s passive/phased automatic update channel; manual updates with `kimi upgrade` are unaffected by this rule.

    How to use it: No additional user-side configuration is required; run the CLI’s normal automatic update check when the update plan is published. Use `kimi upgrade` to skip the wait.

    Tip: For enterprise deployments, check the published update manifest to confirm whether a group with a long delay is actually covered.

Alibaba · Qwen

Qwen4 Max
Full history (673) →

New in this edition · 57

  • AgentQoder

    Qoder enables assigning and reassigning Issues to Agent Teams

    Accounts with project collaboration support can assign and reassign Issues to Agent Teams, and view work that requires human intervention. Work started from a comment retains the initial Issue assignment.

    Usage details

    Who can use it: Qoder 0.5.16; accounts with project collaboration permissions.

    How to use it: Open an Issue; select an Agent Team in the assignment field or reassign it, then start the work from a comment.

    Tip: Separate work requiring human intervention on the team board before execution begins.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • AvailabilityQoder

    Qoder improves device selection and SSH continuity for remote workspace tasks

    New tasks use the selected device, and attachments can be uploaded before sending a message. Tasks accepted by a remote host continue running even if the managed SSH connection drops.

    Usage details

    Who can use it: Qoder 0.5.16; device connections and managed SSH workspaces.

    How to use it: Select the execution device when starting a new task, and upload any attachments before sending the message. If the SSH connection drops, monitor the accepted remote run in the task monitor.

    Tip: For long-running tasks, upload important inputs as attachments before sending to guard against connection interruptions.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • FeatureQoder

    Qoder adds remote access and control for devices on different networks

    Qoder devices signed in to the same account can connect across different networks. When Allow control is enabled on the target device, users can remotely access permitted workspaces and tasks.

    Usage details

    Who can use it: Qoder 0.5.16; devices linked to the same account and users with control permission on the target device.

    How to use it: Update to Qoder 0.5.16; select the target device linked to the same account from the device connections section and enable Allow control on the target device.

    Tip: Narrow the target device's workspace and task permissions before granting remote control.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • CapabilityQoder

    Qoder securely edits password-protected Word and Excel files

    Supported password-protected Word and Excel files can be opened by entering the password. Their encryption is preserved when they are edited and saved.

    Usage details

    Who can use it: Qoder 0.5.16; supported Word and Excel files.

    How to use it: Add the file to a task and enter its password when prompted to open it; after editing, save it from Qoder.

    Tip: Use only the password field in the file-opening prompt instead of writing the password in the task text.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • CapabilityQoder

    Qoder recovers conversations that would not open and plans after a restart

    Fixed an issue where some conversations would not open and plans awaiting approval disappeared from view after a restart. This makes access to task history and pending plans more reliable.

    Usage details

    Who can use it: Qoder 0.5.16; all supported desktop workflows.

    How to use it: Update Qoder to 0.5.16; try reopening the conversation from task history or the plan after restarting.

    Tip: After updating, check the task monitor for a plan that appears to be missing before starting a new task.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • Plugin / IntegrationQoder

    Qoder PPT plugin can embed charts with editable Excel data

    Charts created by the PPT plugin can now embed editable Excel data. This lets users change a chart’s source data after adding it to a presentation.

    Usage details

    Who can use it: Qoder 0.5.16; presentation tasks using the PPT plugin.

    How to use it: Enable the PPT plugin and, when creating a presentation task, ask for the chart to be editable with Excel data. Open the chart and embedded data in the output.

    Tip: Before delivering the presentation, open and verify the Excel data range and chart link.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • CapabilityQoder

    Qoder improves remote Office viewing and Windows/Linux updates

    Fixed save conflicts and display issues after refreshing remote Office files. Also fixed Windows installation/update errors and an issue that blocked subsequent updates for Linux DEB and RPM installations.

    Usage details

    Who can use it: Qoder 0.5.16; remote Office files, and Windows and Linux DEB/RPM packages.

    How to use it: Update to Qoder 0.5.16; refresh the remote Office file after a save conflict, then check for a new update through Help/Update.

    Tip: After updating a Linux package, reopen the application and verify the version.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • SettingQoder

    Qoder adds an output limit setting for supported custom models

    The response output limit can be configured for supported custom models. This lets users limit long responses and control cost and latency.

    Usage details

    Who can use it: Qoder 0.5.16; custom model configurations that support output limits.

    How to use it: Open a supported custom model in model settings, enter the desired upper limit in the output limit field, and save.

    Tip: For code generation, set a limit high enough to meet the task's expected output, rather than the entire file.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • FeatureQoder

    Qoder enables sharing conversation messages and output files

    Completed messages in local or synchronized remote conversations can be selected and saved to the device. Selected messages and output files can be sent to a discussion or Issue in the same project.

    Usage details

    Who can use it: Qoder 0.5.16; local or synchronized remote conversations, and accounts with permission to share in the same project.

    How to use it: Select completed messages; save the output files or use the share action to choose a discussion or Issue in the same project.

    Tip: When transferring content to an Issue, keep the context concise by selecting only the messages needed for the decision and files.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • CapabilityQoder

    Qoder preserves attachments when retrying messages with text and images

    A bug that caused attachments to disappear when retrying a message containing text and images has been fixed. Retry now preserves the original attachments.

    Usage details

    Who can use it: Qoder 0.5.16; task messages with text and image attachments.

    How to use it: Send a message with text and images; if an error occurs, use the Retry action and verify that the attachments remain attached to the message.

    Tip: For large attachments, check the task draft’s attachment list before retrying.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • SettingQoder

    Qoder supports PAC proxy configuration via a URL or local file

    A PAC proxy can be configured for network connections via a URL or local file. Connection issues can be diagnosed from the sign-in screen or Settings.

    Usage details

    Who can use it: Qoder 0.5.16; networks using a PAC file or URL.

    How to use it: Open the network/proxy section in Settings; enter a PAC URL or select a local PAC file. If needed, run connection diagnostics from the sign-in screen.

    Tip: First test a PAC rule that includes only company domains; routing all traffic through the proxy may add unnecessary latency.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • Plugin / IntegrationQoder

    Qoder now lets users share eligible plugins and Skills via links

    Eligible plugins and standalone Skills can be shared via links where public Marketplace sharing is available. Recipients can preview and install the content from the link.

    Usage details

    Who can use it: Qoder 0.5.16; workspaces with public Marketplace sharing enabled and eligible plugin/Skill types.

    How to use it: Open a shareable plugin or Skill in Marketplace, create and send a sharing link. Recipients can open the link to preview the content and select Install.

    Tip: Before sharing, confirm the Skill’s requested permissions and MCP connections with the recipient.

    ✓ Verified by official sourceSource: Qoder Release Notes — Qoder 0.5.16
  • SettingQoderWake

    QoderWake adds a catch-up window for missed scheduled runs

    The Catch up missed runs option has been added to the advanced settings for scheduled tasks. If a device misses a run because it is asleep, the most recent missed run starts automatically if it falls within the specified window.

    Usage details

    Who can use it: QoderWake 1.2.20; scheduled tasks and sleep/offline interruptions.

    How to use it: Open the scheduled task’s Advanced settings, select Catch up missed runs, and set the time window for accepting missed runs.

    Tip: Use a short window on laptops that sleep frequently, and a wider but controlled window on server-like devices.

    ✓ Verified by official sourceSource: QoderWake Release Notes — 1.2.20
  • FeatureQwen Code

    Qwen Code can send a message to another session with /send

    The /send command has been added to Web Shell; a message can be transferred from one session to another. Keeping sent replies visible was also improved in the same nightly release.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; multiple sessions in Web Shell.

    How to use it: Type /send in Web Shell, select the target session, and send your message.

    Tip: Check the target session’s workspace to avoid sending a code change to the wrong project.

  • Plugin / IntegrationQwen Code

    Qwen Code can configure generic outbound webhook delivery

    Generic outbound webhook delivery for managed channels and configuration of a generic webhook target in Web Shell have been added. Events can now be sent to a generic webhook target outside the built-in adapters.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; external systems that accept webhooks.

    How to use it: Open the Web Shell channel settings; enter and save the generic webhook target URL and required delivery settings, then send a test event.

    Tip: In production, use a receiver that verifies signatures and prevents duplicate event IDs.

  • Plugin / IntegrationQwen Code

    Qwen Code adds a GitHub adapter to managed channels

    GitHub has been added to Qwen Code as the third managed channel adapter. This provides a new integration surface for GitHub events and conversations through the managed channel path.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; installations with managed channel permissions and a GitHub connection.

    How to use it: Install the nightly release; select the GitHub adapter in the Web Shell channel manager and complete GitHub authorization.

    Tip: Start with a test repository and limited event scope, then move to a production repository.

  • CapabilityQoderWake

    QoderWake no longer loses the result of commands moved to the background

    A bug that caused a command moved to the background to end its task early without returning the command result has been fixed. When the command completes, its output is now attached to the task result.

    Usage details

    Who can use it: QoderWake 1.2.20; commands moved to the background.

    How to use it: Start a long-running command and move it to the background; when the task completes, open the command output from the result card.

    Tip: Do not run the same command a second time before confirming that it has completed.

    ✓ Verified by official sourceSource: QoderWake Release Notes — 1.2.20
  • CapabilityQoderWake

    QoderWake no longer accidentally reruns background tasks

    A bug that caused a task continuing in the background to be rerun or marked as failed after leaving the conversation page has been fixed. Task status now remains more consistent with background execution.

    Usage details

    Who can use it: QoderWake 1.2.20; conversation tasks moved to the background.

    How to use it: Start the task, leave the conversation, and track a single execution record in the Tasks/monitor view.

    Tip: For long-running tasks, use the execution record in the task monitor as the source of truth for status instead of returning to the page.

    ✓ Verified by official sourceSource: QoderWake Release Notes — 1.2.20
  • CapabilityQoderWake

    QoderWake shows context usage and compacts idle conversations

    The conversation input box shows how much of the context window is in use and opens details when clicked. When an on-device conversation is idle, older messages can be compacted to make room for subsequent work.

    Usage details

    Who can use it: QoderWake 1.2.20; device conversations running this version.

    How to use it: Update QoderWake to 1.2.20; click the context usage indicator in the conversation input box and run the compaction action in an idle conversation.

    Tip: Before starting a long task, check usage details and compact any unnecessary older conversation.

    ✓ Verified by official sourceSource: QoderWake Release Notes — 1.2.20
  • AvailabilityQoderWake

    QoderWake fixes commands failing on some Windows devices

    A problem that caused every command in a conversation to fail on some Windows devices has been fixed. With QoderWake 1.2.20, command execution is usable again on these devices.

    Usage details

    Who can use it: QoderWake 1.2.20; affected Windows devices.

    How to use it: Update QoderWake to 1.2.20 on Windows and test execution with a simple command.

    Tip: On enterprise Windows devices, use a low-risk, read-only command for the initial test.

    ✓ Verified by official sourceSource: QoderWake Release Notes — 1.2.20
  • CapabilityQwen Code

    Qwen Code delivers background async Hook output to the model

    Output from an async Hook completed in the background is now delivered to the model. This prevents new information produced by the Hook from being lost during the same task.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; CLI and Web Shell tasks using Hooks.

    How to use it: Install the nightly build; define a Hook that produces async output and check the transcript to confirm the model uses the output after the Hook completes.

    Tip: Keep Hook output short and structured to reduce noise in the model context.

  • APIQwen Code

    Qwen Code CLI preserves custom command arguments exactly

    The CLI preserves the literal content of custom command arguments. This prevents shell-like characters or formatting in custom commands from being changed unintentionally.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Qwen Code CLI custom commands.

    How to use it: Install the CLI nightly version; run a custom command containing literal characters and check that the argument remains unchanged in the transcript/command output.

    Tip: Still validate command arguments as untrusted shell text; literal preservation does not replace security checks.

  • Plugin / IntegrationQwen Code

    Qwen Code now delivers multipart DingTalk replies without truncation

    Fixed a ledger mismatch that caused long replies sent in chunks to DingTalk to be truncated. Multipart delivery now completes while preserving all chunks.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; DingTalk managed channel connections.

    How to use it: Update the DingTalk channel to the nightly version; generate a long reply and check that all chunks reach your channel.

    Tip: For long replies, also check the channel's chunk limits; complete delivery does not remove those limits.

  • Plugin / IntegrationQwen Code

    Qwen Code deduplicates the DingTalk reconnection chain

    When a DingTalk connection drops, a single disconnect can no longer start multiple reconnect chains. The reconnection flow is now more predictable.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; DingTalk managed channel connections.

    How to use it: Update the DingTalk channel to the nightly version; briefly interrupt the connection and monitor for a single reconnect chain and the return of message flow.

    Tip: Do not manually start multiple reconnection attempts at the same time.

  • AvailabilityQwen Code

    Qwen Code enables idle session suspension for Hosted Harness

    The first phase of Stateless Hosted Harness adds a flow for suspending idle sessions and resuming them when needed. This aims to reduce resource usage for hosted sessions that remain unused for long periods.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Managed Agent sessions using Hosted Harness.

    How to use it: Start a Managed Agent session using Hosted Harness on the nightly release; monitor its suspension status when idle and its readmission status when resumed.

    Tip: For long-running tasks, don't start a new task without checking that the session is ready to resume.

  • AgentQwen Code

    Qwen Code safely recovers lost Harness admission records

    Managed Agent safely recovers missing Harness admission records. This reduces the chance that a resumable session gets stuck because its admission record was lost.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Managed Agent and Hosted Harness.

    How to use it: Use the nightly release; in the session with the missing admission record, follow the Managed Agent recovery notification and run the recommended resume action.

    Tip: Do not manually restart the same session in parallel during recovery; this reduces the risk of duplicate execution.

  • FeatureQwen Code

    Qwen Code keeps background jobs visible in the sidebar

    Session rows in the Web Shell sidebar mark jobs running in the background. Users can more easily see which jobs are still running while they are in another session.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell sidebar.

    How to use it: Move a task to the background; switch to another session and watch the running indicator in the Web Shell sidebar.

    Tip: After moving long-running jobs to the background, check the sidebar so you don't miss when they finish.

  • CapabilityQwen Code

    Qwen Code notifies users when Managed Agent attachments expire and when suspension delays occur

    Managed Agent notifies users when attachments expire and tracks suspension delays. Users are visibly informed that an attachment is no longer valid instead of encountering a silent failure.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Managed Agent attachments and suspended sessions.

    How to use it: Use the nightly build; open the notification shown when an attachment expires or a suspension delay occurs in a Managed Agent session with an attachment.

    Tip: Check the session's current context before re-uploading the expired attachment and restarting the task.

  • CapabilityQwen Code

    Qwen Code aligns Managed Agent errors with DAC and recovers

    Managed error states in Web Shell are now aligned with DAC, and the recovery flow has been made more stable. The recovery path a session follows after an error is now more consistent.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Managed Agent/Web Shell.

    How to use it: Use the nightly release; when an error occurs in a managed task, follow the recovery action on the error card and reopen the session.

    Tip: Do not start a new recovery before saving the session ID shown on the error card.

  • APIQwen Code

    Qwen Code brings Mods execution to ACP sessions

    Mods can now be run in ACP sessions. This allows customized terminal behavior in programmatically driven ACP flows.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; ACP clients and sessions using Mods.

    How to use it: Update the nightly build and your ACP client; define an active Mod in an ACP session and start the session.

    Tip: Keep the Mod's file and tool permissions limited to what is needed before making it available to ACP.

  • CapabilityQwen Code

    Qwen Code automatic jobs no longer block worktree restoration

    Fixed an issue where the server’s automatic job blocked worktree restoration. Background jobs are now handled more reliably during session or branch restoration.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell/serve workflows that use worktrees.

    How to use it: Install the Nightly version; in a session that uses a worktree, restore or switch branches and monitor the ongoing automatic job.

    Tip: Save the working tree state of critical files before restoring.

  • APIQwen Code

    Qwen Code Python SDK separates initialization from caller cancellation

    The Python SDK initialization process is now isolated from the caller's cancellation signal. This reduces the risk of SDK initialization being disrupted—and subsequent calls being affected—even when a higher-level task is canceled.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Python SDK integrations.

    How to use it: Update the Python SDK to the nightly release; cancel the caller's task during initialization and test that the SDK closes its resources safely.

    Tip: Manage initialization and task cancellation with separate timeouts.

  • APIQwen Code

    Qwen Code fixes case-collision issues in SDK caller headers

    The SDK now applies HTTP headers added by the caller without creating case-sensitive duplicate entries. This reduces API incompatibilities caused by sending the same header twice.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; integrations using the SDK.

    How to use it: Update the SDK to the nightly release; rerun a call that adds the same HTTP header with different capitalization.

    Tip: Use a single canonical format for header names in client code to reduce provider differences.

  • APIQwen Code

    Qwen Code SDK stops delivering buffered SSE data after cancellation

    The SDK no longer continues sending buffered SSE data when a request is canceled. Late chunks from a canceled stream no longer mix with a new session or callback.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; integrations using the SDK and SSE.

    How to use it: Update the SDK; start an SSE request, then cancel it and verify that no new data callbacks arrive after cancellation.

    Tip: Also close the consumer queue in the cancellation callback; do not rely on SDK cancellation alone.

  • APIQwen Code

    Qwen Code SDK clears the permission deadline when a callback fails

    The SDK's permission deadline is now cleared when the permission callback fails. This prevents subsequent permission requests from being affected by the timeout of a previous failed call.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; SDK permission callbacks.

    How to use it: Update the SDK to the nightly release; test the permission callback so it returns an error, then verify that the next permission request uses a new deadline.

    Tip: Also cancel any pending timeout in your own application when a callback fails.

  • APIQwen Code

    Qwen Code SDK enforces the peer frame limit before dispatch

    The SDK enforces the peer frame size limit before dispatch. Oversized frames are rejected before reaching the call layer, helping preserve connection stability.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; integrations using the SDK/WebSocket.

    How to use it: Update the SDK; send a request that exceeds the frame limit and verify that it is rejected early and clearly.

    Tip: Measure frame size on the application side in advance to avoid unnecessary network calls.

  • AvailabilityQwen Code

    Qwen Code strengthens SSE stream capacity and cleanup of stalled writers

    The serve layer unifies admission and occupancy calculations across all SSE surfaces, while also preserving listener capacity and cleaning up stalled writers. This reduces unnecessary connection lockups in concurrent streams.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; qwen serve and clients using SSE.

    How to use it: Install the nightly release; start qwen serve, open concurrent SSE streams, and monitor the capacity/status indicators.

    Tip: Add reconnection and idempotent event handling to SSE clients in production.

  • CapabilityQwen Code

    Qwen Code strengthens verify-pr monitoring and VM checks

    Watch-state format validation, container nonce handling, and virtual machine descriptions have been strengthened in the verify-pr flow. This reduces the risk of invalid states being accepted incorrectly and the monitoring flow becoming unstable.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; repositories that use verify-pr.

    How to use it: Install the nightly release; run /verify-pr on a PR and check the watch state and container/VM summary on the results screen.

    Tip: Before automated publishing, review the watch-state and environment summary in the verify-pr output.

  • FeatureQwen Code

    Qwen Code adds a unified chat surface and capability safeguards to Web Shell

    A unified chat surface contract and capability guards that check which capabilities the surface supports have been added to Web Shell. This lets the same chat flow identify unsupported actions earlier and more clearly.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell.

    How to use it: Install the nightly release and open a chat in Web Shell; for an unsupported action, note the capability warning shown by the surface.

    Tip: Before using a feature across different Qwen Code surfaces, first check that surface’s capability indicator.

  • AvailabilityQwen Code

    Qwen Code adds a Traditional Chinese interface to Web Shell

    Traditional Chinese localization has been added to the Web Shell interface catalogs. This makes interface text available in Traditional Chinese.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell and the supported language selector.

    How to use it: Open the nightly release; select Traditional Chinese from Web Shell Settings or the language selector.

    Tip: If localization is incomplete, switch languages in the same session and reload to refresh the catalog cache.

  • CapabilityQwen Code

    Qwen Code managed agents discover and pin default workspace Skills

    Managed Agent can discover and pin default workspace Skills. This helps sessions start more consistently with the workspace's expected Skill set.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Managed Agent workspaces.

    How to use it: Install the nightly version; refresh Managed Agent workspace settings and pin the discovered default Skill list.

    Tip: Review the pinned Skill list after upgrades to prevent outdated instructions from carrying over.

  • CapabilityQwen Code

    Qwen Code marks incoming message truncation in managed channels

    Cases where an incoming message is truncated in managed adapters are now marked. This makes it clear when a response is based on truncated content rather than the full message.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; managed channel adapters.

    How to use it: Install the nightly release; send a long channel message and check for the inbound truncation marker in the session.

    Tip: When you see a truncation marker, resend critical details in a short follow-up message.

  • CapabilityQwen Code

    Qwen Code distinguishes combined shell output from execution errors

    Shell output is labeled separately from execution errors as a combined stdout/stderr stream. This reduces the chance that the model will confuse normal command output with an error signal.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; tasks that use the shell.

    How to use it: Install the nightly release; run a shell command that produces stdout and stderr, then check that the stream and execution error labels are separate in the tool card.

    Tip: In error-handling code, evaluate the output type label as well as the exit code.

  • CapabilityQwen Code

    Qwen Code CLI preserves pending cleanup when a handler is removed

    Fixed an issue where pending cleanup work was lost when handlers were removed in the CLI. Session shutdown and resource cleanup now complete more consistently.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Qwen Code CLI.

    How to use it: Update the CLI to the nightly version; stop a session, then check that temporary resources are cleaned up after the handler/extension removal flow.

    Tip: Allow a few seconds for the normal exit flow before forcibly terminating the session.

  • CapabilityQwen Code

    Qwen Code falls back to the session's starting commit for the change overview

    The commit fallback for the Web Shell change overview is now limited to the session's starting commit. This prevents the overview from being calculated against an incorrect remote or later commit.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell change overview.

    How to use it: Install the nightly build; make changes in a session, open the change overview, and check the diff against the starting commit.

    Tip: Also compare the starting commit with the current working tree to verify the overview's scope.

  • FeatureQwen Code

    Qwen Code adds a bottom-sheet model picker for touchscreens

    The Web Shell model picker can open as a bottom sheet on touchscreens. This makes switching models more usable on smaller screens.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell sessions on touchscreens.

    How to use it: Open Web Shell on a touch device; tap the model picker, then select and confirm a model in the bottom sheet.

    Tip: Before making a selection on a small screen, check that the current model and reasoning settings are preserved.

  • FeatureQwen Code

    Qwen Code Generated UI cards save view state to the host

    Generated UI cards can store view state through the host. A card's selection or view preference can be retained when it is reopened.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Generated UI cards and host-supported surfaces.

    How to use it: Install the nightly version; change the view or selection state on a Generated UI card, close the session, and reopen the card.

    Tip: Save the default view for frequently reused panels to reduce the need to recreate the same layout in every session.

  • FeatureQwen Code

    Qwen Code marks Goal sessions in the sidebar

    The Web Shell sidebar displays a separate marker for sessions that include Goal. This makes it easier to distinguish goal-oriented executions from regular chats.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell and Goal sessions.

    How to use it: Start a session with Goal; find it in the Web Shell sidebar using the Goal marker on the session row.

    Tip: Keep long-running Goal sessions in a separate filter or workspace so they aren't confused with regular chats.

  • CapabilityQwen Code

    Qwen Code narrows Grep fallback strategies with Git awareness

    Grep fallback strategies are now limited to a Git-aware corpus. This reduces searches extending to irrelevant sources and prevents false matches.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; CLI and Web Shell tasks that use Grep.

    How to use it: Install the nightly release; run a Grep search in a Git workspace and check that the results remain limited to the repository corpus.

    Tip: If search scope matters, specify the working directory explicitly and compare the result with git status.

  • CapabilityQwen Code

    Qwen Code stops canceled batch image reads

    A bug that let subsequent image-reading jobs continue after a batch was canceled has been fixed. This reduces unnecessary image processing and delays after cancellation.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell batch image-reading flows.

    How to use it: Install the nightly build; start a task with multiple images, cancel the batch, and check that no new image-reading calls are made.

    Tip: Cancel large image batches early to reduce unnecessary token and processing usage.

  • APIQwen Code

    Qwen Code Java client preserves line breaks in file content responses

    The Java client now returns line breaks in file content responses. Multiline files are no longer converted to a single line or malformed in Java integrations.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Java SDK/client integrations.

    How to use it: Update the Java client to the nightly release; make a multiline file-read call and verify the line breaks in the returned content.

    Tip: For file comparisons, also normalize platform-specific line-ending conversions.

  • CapabilityQwen Code

    Qwen Code preserves fence-like content when copying code and math

    Fixed a copying issue that corrupted fence-like content when extracting code and math. Nested markers in replies are now preserved more accurately when copied.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell and supported reply-copying flows.

    How to use it: Install the nightly version; select a reply containing nested code fences or math markers and use the Copy action.

    Tip: Before pasting into a document, briefly preview the copied text to check the fence closures.

  • FeatureQwen Code

    Qwen Code shows artifact counts from the session list on hover

    Web Shell shows artifact counts from session-list metadata on hover. This lets users see how many outputs a session contains without opening it.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell session list.

    How to use it: Hover over or click a session with an artifact indicator in the Web Shell session list; check the count before opening the session.

    Tip: Prioritize sessions that produce many outputs based on their artifact counts.

  • CapabilityQwen Code

    Qwen Code increases the document rendering budget for recent transcript copies

    The document renderer budget in web templates has been increased when copying recent transcripts. Long transcripts are now less likely to be truncated when copied.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell transcript copying.

    How to use it: Install the nightly release; copy a long transcript through the Copy/export flow and check that the final section is included in the output.

    Tip: For very long histories, still limit the output size by selecting only the necessary turn.

  • CapabilityQwen Code

    Qwen Code limits context usage for follow-up suggestions

    The context used to generate follow-up suggestions is now limited. This reduces unnecessary token consumption when generating suggestions for long histories.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; surfaces with follow-up suggestions enabled.

    How to use it: Install the nightly version; wait for a follow-up suggestion to appear in an empty input field, then observe token usage in a session with a long history.

    Tip: Before using a follow-up suggestion, narrow the suggested text with a brief request to reduce unnecessary context.

  • AvailabilityQwen Code

    Qwen Code provides verify-pr watch links before sign-in

    Web Shell can now provide PR watch deep links before authentication. When users open a link, they can first be routed to the relevant watch path and then complete the sign-in flow.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell PR watch links.

    How to use it: Open a PR watch deep link; if prompted to sign in, complete authentication and return to the watch target of the link.

    Tip: Before signing in, verify in the URL that the deep link you plan to share with your team points to the correct repo and PR.

  • CapabilityQwen Code

    Qwen Code retries loading memory content after a refresh

    Web Shell retries loading memory content that could not be loaded after a refresh. This reduces the chance that the memory section remains empty during temporary refresh interruptions.

    Usage details

    Who can use it: Qwen Code v0.28.0-nightly.20261118; Web Shell memory view.

    How to use it: Install the nightly release; refresh the memory view and observe the reload behavior if a temporary read error occurs.

    Tip: Do not start a new forget or edit operation before the memory content appears.

Watchlist

Muse Wonder Full

50
Meta · Meta AI

Tied with Gemini 4 Argon at 50 points, but just outside the TOP 10 because it ranks 11th in the Artificial Analysis table. Meta’s strongest Muse model; no public API, available to Muse Code users.

Kimi K3.1 Pro

46
Moonshot AI · Kimi

12th place; still the strongest open-weight model. API $2 / $10, weights on Hugging Face, also available on Amazon Bedrock. Kimi Code 2.7.0 added an Amazon Bedrock provider and a full Kimi Code server on SSH machines; Kimi Code Desktop 1.4.0 can connect to Claude Code, Codex and Cursor sessions.

Muse Wonder

44
Meta · Meta AI

Meta’s standard Muse Wonder model; 13th place with 44 points. In preview on Google Agent Platform as `muse-wonder-1.0`.

Qwen4 Max

43
Alibaba · Qwen

Released on November 16; 14th place with 43 points. Available in Model Studio as `qwen4-max` (1M context, Singapore pricing $1,50 / $4,50), QwenCloud Token Plan, QwenWork, QwenStudio and OpenRouter (`qwen/qwen4-max`). Selectable in Qoder from day one. Not open-weight.

Claude Haiku 5.5

36
Anthropic · Claude

Anthropic’s fast, inexpensive 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, `muse-ember-1.0` in Meta Model API and available on all Muse Code plans.

Sources