Kimi K3.1 Pro weights released, ChatGPT Images 2.6 arrives

New in this edition · 9

• 04 Nov
  AvailabilityGemini

  UCP-powered direct checkout expands to Canada in the Gemini app

  Direct checkout powered by Universal Commerce Protocol (UCP) is beginning to roll out to users in Canada in Google Search AI Mode and the Gemini app. For eligible listings, users can buy directly from select Canadian retailers using Google Pay.

  Usage details

  Who can use it: Canada; Gemini app and Search AI Mode; initially, eligible retailers such as lululemon, select Shopify merchants, Shoppers Drug Mart, and Simons.

  How to use it: Search for an eligible product or retailer in the Gemini app or Search AI Mode; select the direct checkout flow shown in the results and complete the purchase using payment and delivery details saved in Google Pay.

  Tip: Verify your payment and shipping details in Google Wallet in advance; eligibility varies by retailer and product.

  ✓ Verified by official sourceSource: Google Canada — Kanada’da agentic shopping ve UCP destekli checkout
• 04 Nov
  APIGemini Enterprise Agent Platform

  Unified gRPC and HTTP streaming arrives for AlphaGenome

  AlphaGenome (alphagenome-004) now supports unified HTTP/2 gRPC streaming and chunked binary protobuf HTTP streaming on Gemini Enterprise Agent Platform. The update enables lower-latency 1 MB track predictions across seven DnaModelService methods, concurrent multi-variant evaluation, and model fold selection; existing `alphagenome-003` HTTP deployments remain supported.

  Usage details

  Who can use it: Gemini Enterprise Agent Platform; AlphaGenome access requires an allowlist, and self-deployed endpoints require suitable Google Cloud GPU capacity.

  How to use it: On your AlphaGenome endpoint, use `PredictionService.StreamDirectRawPredict` for gRPC (shared endpoint, port 8081) or `streamRawPredict` for HTTP (dedicated endpoint, port 8080); use `alphagenome_cloud_sdk` to configure `ALL_FOLDS` or select a specific fold.

  Tip: For large variant scans, use concurrent `score_ism_variants` calls and gRPC binary streaming to reduce the risk of HTTP gateway timeouts.

  ✓ Verified by official sourceModels: AlphaGenome (alphagenome-004), alphagenome-003Source: Google Cloud sürüm notları — 4 Kasım 2026Source: AlphaGenome — Gemini Enterprise Agent Platform dokümantasyonu
• 04 Nov
  APIGemini API

  Voice remixing API and AI Studio arrive for Gemini 3.8 Flash TTS

  You can modify an existing prebuilt, Extended Voice Library, or designed voice with natural-language instructions to create a standalone, stored `voice_...` ID. The feature is available interactively in Google AI Studio and through `POST /v1beta/voices`; the generated `voice_...` ID can be used for synthesis with either of the two Gemini 3.8 TTS models.

  Usage details

  Who can use it: Projects on Gemini API Paid Tier 1 or above and Google AI Studio; project quota of 200 stateful voices per created voice, with storage for 1 year after last use.

  How to use it: In Google AI Studio, open Generate Speech > Voice remixing; for the API, call `POST /v1beta/voices` with the fields `store: true`, `type: "prompted"`, `prompted.base_voice`, and `prompted.input`, then use the returned `voice_...` ID in a TTS request.

  Tip: Change only one or two attributes (for example, pitch and pacing) with each remix; listen to the returned `sample_audio` preview before proceeding to the next remix.

  ✓ Verified by official sourceModels: gemini-3.8-flash-tts, gemini-3.8-flash-lite-ttsSource: Google AI Studio değişiklik günlüğü — 4 Kasım 2026Source: Voice remixing kılavuzu
• 04 Nov
  Plugin / IntegrationGemini

  monday.com becomes available as a Connected App in Gemini

  monday.com is now available as a Connected App in Gemini for personal, work, and school Google accounts; users can ask about information on their boards in a Gemini chat using `@monday.com`. On Google Workspace, the integration is also available through Marketplace for working with monday data in Gmail, Docs, Slides, and Sheets.

  Usage details

  Who can use it: Gemini web app, mobile app, and Windows app; English; countries and regions where Gemini and monday.com are supported; personal, work, or school Google account.

  How to use it: Connect monday.com from Gemini Settings > Connected Apps, or type `@monday.com` in any Gemini chat and select the app; Workspace users can also install the monday.com integration from Google Workspace Marketplace.

  Tip: Narrow your first query to a specific board and time range (for example, overdue tasks); check the connected monday workspace’s access scope before using the response.

  ✓ Verified by official sourceSource: monday.com — Google Gemini’de kullanılabilir olduSource: Google Gemini Connected Apps uygunluk tablosu
• 04 Nov
  OtherGemini CLI

  Gemini CLI nightly fixes extension, model, hook, and sandbox bugs

  The November 4 nightly release fixes streaming errors in extension downloads, preserves an explicitly selected `gemini-2.5-flash` model during dynamic model resolution, and fixes `SubagentStop` hook matching. It also improves timely replay of ACP session/load history and `--userns=keep-id` support in rootless Podman.

  Usage details

  Who can use it: Gemini CLI pre-release nightly channel; release tag `v0.67.0-nightly.20261104.g89e9d25d6`.

  How to use it: Upgrade your nightly installation to `v0.67.0-nightly.20261104.g89e9d25d6`; then rerun extension release download, hook, ACP session/load, and rootless Podman workflows.

  Tip: After upgrading the nightly build, teams using CI/CD or Podman sandbox should run a brief smoke test to confirm that previous model selections and hook configurations are preserved.

  ✓ Verified by official sourceModels: gemini-2.5-flashSource: Gemini CLI nightly 4 Kasım 2026 sürüm notu
• 04 Nov
  FeatureGemini Enterprise

  Switching between teams in the Gemini Enterprise mobile app reaches GA

  The Gemini Enterprise mobile app can now switch between multiple teams without signing out and scanning a QR code again. The first team opened is saved to the list; users can rename or remove a team from that list.

  Usage details

  Who can use it: Gemini Enterprise mobile app; documented as a GA feature for Standard, Plus, Pay-as-you-go, and Frontline editions.

  How to use it: Sign in to the Google Gemini Enterprise mobile app with your company account; after the initial connection, select another team from the teams list, or rename or remove a team using the list item’s options.

  Tip: Administrators working across multiple organizational environments can reduce QR code and sign-in steps by saving each team with a meaningful name.

  ✓ Verified by official sourceSource: Google Cloud sürüm notları — 4 Kasım 2026Source: Gemini Enterprise mobil uygulamasını kullanma
• 03 Nov
  APIGoogle Cloud CLI (Agent Registry)

  `--scanning-state` removed from Agent Registry CLI

  With Google Cloud CLI 592.0.0, the `--scanning-state` flag was removed from the Agent Registry skills create and update commands. Use `--scanning-policy` instead.

  Usage details

  Who can use it: Google Cloud CLI 592.0.0 and later; Agent Registry skill creation or update workflows.

  How to use it: Use the `--scanning-policy` flag instead of `--scanning-state` in the `gcloud agent-registry skills create` and `gcloud agent-registry skills update` commands.

  Tip: Update the flag in CI/CD scripts along with the version upgrade; scan your command templates and automation repositories for the old flag.

  ✓ Verified by official sourceSource: Google Cloud sürüm notları — Cloud SDK 592.0.0
• 03 Nov
  AvailabilitySnowflake AI Functions

  Gemini 4 Argon reaches general availability in Snowflake AI Functions

  According to Snowflake’s official release notes, the `gemini-4-argon` model is GA in AI Functions and has been added to regional and cross-region usage options. Since I could not find separate confirmation of this Snowflake access through Google’s own channels, the entry is marked `verified=false`.

  Usage details

  Who can use it: Snowflake AI Functions; regional and cross-region availability depends on Snowflake’s Cortex AI Functions model table.

  How to use it: Select `gemini-4-argon` as the model in Snowflake AI Functions and check the relevant regional or cross-region support table before making a call.

  Tip: Before going to production, verify model access, quota, and pricing separately for the region you selected in Snowflake.

  ! UnverifiedModels: Gemini 4 ArgonSource: Snowflake sürüm notu — 3 Kasım 2026: Google Gemini 4 Argon GA
• 03 Nov
  AvailabilityGoogle Vids

  The rollout of creating custom avatars from images in Google Vids has been paused

  The rollout of the feature for creating custom avatars in Google Vids by uploading an image from a local computer, Drive, or Photos was paused on November 3. Google did not provide a restart date and updated the feature’s rollout timeline to TBD.

  Usage details

  Who can use it: The rollout of the feature to the Business, Enterprise, Education Plus, personal Google accounts, and relevant add-on plans listed in the announcement is paused.

  How to use it: You can’t enable this feature right now; when the rollout resumes, open the avatar creation flow in Google Vids and use the image upload option.

  Tip: Plan team training and content templates after the rollout resumes; in the meantime, use Vids’ existing flow for creating avatars from scratch.

  ✓ Verified by official sourceSource: Google Workspace Updates — Google Vids özel avatar güncellemesi

New in this edition · 15

• 04 Nov
  AgentClaude Marketplace

  AirOps added to Claude Marketplace

  AirOps offers an agent-based platform for AI search strategy, content, and multichannel execution. It helps marketing teams connect brand performance in AI search with revenue and pipeline outcomes.

  Usage details

  Who can use it: Partner agent/product listed on Claude Marketplace; enterprise purchasing and an AirOps account are required.

  How to use it: Open the AirOps listing on Claude Marketplace and submit a purchase request; connect your brand, content, and multichannel campaign goals during partner setup.

  Tip: For your first pilot, pair AI search visibility with a measurable pipeline metric.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  AgentClaude Marketplace

  Asana AI Teammates added to Claude Marketplace

  Asana AI Teammates offers ready-to-use agents that use Asana Work Graph context and have separate identities, comprehensive permissions, and audit trails. They can route decision points in multistep work to the responsible person.

  Usage details

  Who can use it: Partner agent listed on Claude Marketplace; an Asana workspace, enterprise purchasing, and permission configuration are required.

  How to use it: Open the Asana AI Teammates listing on Claude Marketplace and submit a purchase request; configure the agent identity and the scope of Asana it can access with limited permissions.

  Tip: Keep the audit trail readable by defining a single workflow and a narrow permission scope for each teammate.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Chainguard Harness added to Claude Marketplace

  Chainguard Harness builds applications from source in isolated microVMs and probes running applications for exploitable vulnerabilities. It provides custom remediation recommendations for Claude’s findings.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; purchasing and access depend on the eligibility of the relevant Enterprise customer.

  How to use it: Open the Chainguard Harness listing on Claude Marketplace and submit a purchase request, then define the application repository and test environment scope through partner setup.

  Tip: Run your first trial on an isolated test application rather than in production, and verify recommended remediations by rescanning.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Checkmarx One added to Claude Marketplace

  Checkmarx One combines rule-based analysis with AI reasoning to find code vulnerabilities. It helps prioritize findings and fix them before code is released.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; enterprise purchasing and access terms vary by listing.

  How to use it: Open the Checkmarx One listing on Claude Marketplace, submit a purchase request, and add the connection to your existing Checkmarx code analysis workflow using the partner guidelines.

  Tip: Use rule-based findings alongside Claude’s prioritization and create a separate remediation queue for high-impact vulnerabilities.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  AgentClaude Marketplace

  Cyera Agent Guardian added to Claude Marketplace

  Cyera Agent Guardian inventories Claude deployments and manages permissions before Claude runs. It helps enforce policies for prompts, files, and tool calls, and test these controls with red-team exercises.

  Usage details

  Who can use it: Partner agent listed on Claude Marketplace; enterprise access and purchasing terms depend on the listing.

  How to use it: Open the Cyera Agent Guardian listing on Claude Marketplace and submit a purchase request; define your deployments and the prompt, file, and tool policies to enforce during partner setup.

  Tip: Before going live, run red-team scenarios with the most restrictive permission policies and measure false blocks.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Datadog Agent Observability added to Claude Marketplace

  Datadog Agent Observability traces agent executions across tools and services. It surfaces production patterns and provides recommendations to improve agent quality, performance, and cost efficiency.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; enterprise purchasing and Datadog access are required.

  How to use it: Open the Datadog Agent Observability listing on Claude Marketplace and submit a purchase request; connect Claude-based agent executions to the partner’s observability workflow.

  Tip: Track success rate, latency, and cost per tool together on your first dashboard; don’t look only at total request volume.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Glean added to Claude Marketplace

  Glean brings together company knowledge and workflows across connected systems. It helps employees find answers from permitted sources and complete multistep work in connected applications.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; connected application permissions and enterprise purchasing terms apply.

  How to use it: Open the Glean listing on Claude Marketplace and submit a purchase request; select the enterprise applications to connect and user access in partner administration.

  Tip: For initial use, verify source permissions with a small user group and leave sensitive repositories out by default.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  AgentClaude Marketplace

  monday agents added to Claude Marketplace

  monday agents uses context from projects, boards, documents, and meetings, along with owners and due dates. It gives PMO, IT, marketing, finance, and operations teams a way to control AI access, approvals, and spending.

  Usage details

  Who can use it: Partner agents listed on Claude Marketplace; requires a monday workspace and enterprise permissions configuration.

  How to use it: Open the monday agents listing on Claude Marketplace and submit a purchase request; define the boards to be used, ownership rules, and approval workflows on the partner side.

  Tip: Initially keep agents’ spending and approval permissions in read-only or human-approval mode.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Netlify Enterprise added to Claude Marketplace

  Netlify Enterprise deploys applications, sites, and agents built with Claude to its global edge network. It provides preview URLs for every change, along with SSO, role-based access, and audit logs.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; Enterprise governance and purchase terms apply.

  How to use it: Open the Netlify Enterprise listing on Claude Marketplace and submit a purchase request; connect a Claude output to a Netlify project to start a preview deployment.

  Tip: Review every change at its preview URL before sending it to production, and keep publishing permissions limited.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  AgentClaude Marketplace

  Rubrik Code Guardian added to Claude Marketplace

  Rubrik Code Guardian examines a secure copy of GitHub or Azure DevOps repositories for multi-step attack paths. It ranks issues by severity and helps turn verified critical findings into tracked tickets.

  Usage details

  Who can use it: Partner agent listed on Claude Marketplace; subject to repository access, purchase, and enterprise authorization terms.

  How to use it: Open the Rubrik Code Guardian listing on Claude Marketplace, submit a purchase request, and select the GitHub or Azure DevOps repositories to scan in the partner connection flow.

  Tip: First scan a read-only repository copy, then enable the ticketing rule for critical findings with team approval.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Runway added to Claude Marketplace

  Runway creates and edits video, images, and audio using natural language. It helps creative and marketing teams turn a brief into finished media assets.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; access and purchase terms depend on Runway and the enterprise agreement.

  How to use it: Open the Runway listing on Claude Marketplace and submit a purchase request; then provide a creative brief in natural language to start a video, image, or audio generation workflow.

  Tip: Include the target format, duration, brand constraints, and revision criteria in the brief from the start.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Pricing / PlanClaude Marketplace

  Committed Anthropic spend can be used for Marketplace solutions

  Eligible Enterprise customers can direct part of their committed Anthropic spend to partner solutions on Claude Marketplace. This makes it possible to purchase Claude and partner products under a single commitment.

  Usage details

  Who can use it: For solutions on Claude Marketplace; eligible Enterprise customers. Eligibility and commitment scope are determined by Anthropic.

  How to use it: Open the Agents & products section on Claude Marketplace, select a listing, and submit a purchase request; confirm with your Anthropic account team whether your committed balance can be applied.

  Tip: Before ordering, confirm in writing which partner products are covered by your commitment.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Tanium Cloud added to Claude Marketplace

  Tanium Cloud provides real-time visibility and control over endpoints. It helps teams find and fix vulnerabilities, misconfigurations, and missing patches.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; access and purchase depend on the relevant enterprise agreement.

  How to use it: Open the Tanium Cloud listing on Claude Marketplace and submit a purchase request; configure the endpoint inventory and patch management scope in the partner account.

  Tip: First monitor critical asset groups under a separate policy, then gradually expand the scope of automated remediation.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  Plugin / IntegrationClaude Marketplace

  Webflow added to Claude Marketplace

  Webflow provides built-in tools for creating, managing, and optimizing websites. It allows marketers, designers, developers, and agents to make changes with governance controls.

  Usage details

  Who can use it: Partner product listed on Claude Marketplace; enterprise access and purchase terms apply.

  How to use it: Open the Webflow listing on Claude Marketplace and submit a purchase request; add your workspace to the Claude workflow using the partner connection.

  Tip: Test change approvals and role limits in a draft workspace before granting publishing permissions.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler
• 04 Nov
  AgentClaude Marketplace

  Zoom Agentic Search added to Claude Marketplace

  Zoom Agentic Search searches meeting summaries, workplace conversations, documents, and connected third-party sources from within Claude. It provides permission-aware results and works without analyzing raw transcripts.

  Usage details

  Who can use it: Partner agent listed on Claude Marketplace; Zoom and connected-source permissions, as well as enterprise purchase terms, apply.

  How to use it: Open the Zoom Agentic Search listing on Claude Marketplace and submit a purchase request; select the Zoom workspace and permitted third-party sources in the partner connection.

  Tip: Before using results for critical decisions, verify the permissions and freshness of the source meeting or document.

  ✓ Verified by official sourceSource: Claude Marketplace’te güvenlik, üretkenlik ve yaratıcı çalışma için yeni ajanlar ve ürünler

New in this edition · 6

• 04 Nov
  ModelChatGPT Images

  ChatGPT Images 2.6 brings faster generation and more consistent editing

  OpenAI made ChatGPT Images 2.6 available on November 4. The model generates images faster than the previous version while better preserving elements such as characters’ clothing and product details during edits; OpenAI says generation is 50% faster.

  Usage details

  Who can use it: Rolling out worldwide across all subscription tiers in ChatGPT and Codex on web, desktop, and mobile.

  How to use it: In ChatGPT, ask for an image in a conversation or open Images via More > Images; upload an existing image and describe the edits you want. You can also use the image generation and editing workflow in Codex.

  Tip: Explicitly lock in character, product, and brand details in your editing prompt; asking for one major change per turn can improve consistency.

  ✓ Verified by official sourceModels: Images 2.6Source: ChatGPT sürüm notlarıSource: Images 2.6 tanıtımıSource: Images in ChatGPT yardım sayfası
• 04 Nov
  AvailabilityChatGPT

  Dots rolling out for general availability in Enterprise, Edu, and Healthcare workspaces

  Dots is becoming generally available in eligible ChatGPT Enterprise, Edu, and Healthcare workspaces; the administrator preview continues through November 18, 2026, and access is off by default during this period. Dots supports data residency and Enterprise Key Management with EU/US inference residency, and may fall under HIPAA in eligible workspaces; workspaces with UAE data or inference residency are not supported.

  Usage details

  Who can use it: Eligible Enterprise, Edu, and Healthcare workspaces; administrator preview from 2026-11-04–2026-11-18, followed by default enablement unless the administrator opts out. Excludes workspaces with UAE data or inference residency.

  How to use it: As an administrator, open Workspace settings > Permissions & roles > Permissions > Workspace default > Workspace capabilities, review Use dots and the relevant computer capabilities, and save. If needed, keep on-request approval and auto-review enabled under Admin Console > Agent Security > Global > Requirements. Members can then create their own dot through the onboarding flow in the mobile app, desktop app, or desktop web.

  Tip: Since access is off during the preview, first enable it only for the roles and members who need it; regularly review personal dots’ connections, usage, and administrator activity.

  ✓ Verified by official sourceModels: GPT-6 AstraSource: ChatGPT Enterprise ve Edu sürüm notlarıSource: ChatGPT çalışma alanlarında dots yönetimiSource: Dot ile çalışmaya başlama
• 04 Nov
  APIOpenAI Images API

  GPT-Image-2.6, complexity, SVG, GIF, and markup editing come to the Images API

  OpenAI released the GPT-Image-2.6 Sunburst and Flare models in its image generation and editing API. The Image API now offers complexity control, SVG and animated GIF output; markup-guided editing has also been added to the supported GPT Image 2.5 and 2.6 models.

  Usage details

  Who can use it: For organizations using the API, on the `v1/images/generations` and `v1/images/edits` endpoints; rate limits vary by usage tier, and organization verification may be required.

  How to use it: Set the `model` field to `gpt-image-2.6-sunburst` or `gpt-image-2.6-flare` in a `POST /v1/images/generations` or `/v1/images/edits` request. Choose `low`, `medium`, or `high` for the `complexity` field; use `output_format: "svg"` or `output_format: "gif"` when you need SVG or GIF output. For GIF generation, you can also set the `frame_count` and `frame_duration_ms` fields.

  Tip: Try `complexity: "low"` for focused edits and `medium` or `high` for scenes with dense text or many objects; higher complexity may increase latency.

  ✓ Verified by official sourceModels: gpt-image-2.6-sunburst, gpt-image-2.6-flareSource: OpenAI API değişiklik günlüğüSource: GPT-Image-2.6 Sunburst model sayfasıSource: GPT-Image-2.6 Flare model sayfası
• 04 Nov
  FeatureChatGPT Images

  ChatGPT Images reuses saved reference photos in later chats

  The new Saved reference images feature lets you name a photo and use it again in later chats without re-uploading it. For example, you can save your dog as Pepper and later request a watercolor portrait of @Pepper.

  Usage details

  Who can use it: Part of the ChatGPT Images 2.6 rollout; access depends on the account and phased rollout.

  How to use it: Upload a photo in the Images workflow and save it under a name, such as Pepper. In a later image prompt, type @Pepper to use the same reference without re-uploading it.

  Tip: Choose a single well-lit reference for the same character or product and give it a distinctive name that will be easy to find later.

  ✓ Verified by official sourceModels: Images 2.6Source: ChatGPT sürüm notları
• 04 Nov
  FeatureChatGPT

  ChatGPT starts generating looping GIFs and adjusting playback speed

  With Images 2.6, ChatGPT can create looping GIFs from prompts; you can also upload a sketch and ask it to animate the character. You can adjust the generated GIF’s playback speed and share it.

  Usage details

  Who can use it: Rolling out to all plans on web, desktop, and mobile as part of the Images 2.6 rollout in ChatGPT.

  How to use it: Ask ChatGPT to create a looping animation, or upload a sketch and ask it to animate the character. Once the result is generated, adjust the playback speed and share it with the Share option.

  Tip: Explicitly describe the starting and ending poses, the sequence of movements, and that the loop should be seamless.

  ✓ Verified by official sourceModels: Images 2.6Source: ChatGPT sürüm notlarıSource: Images 2.6 tanıtımı
• 04 Nov
  AvailabilityDaybreak Red

  Daybreak Red becomes available to approved customers through AWS GovCloud (US-East)

  OpenAI expanded access to GPT-5.6 Cyber to the AWS GovCloud (US-East) environment through Daybreak Red. Authorized security teams can conduct vulnerability research, exploit validation, and security testing through Amazon Bedrock; access is subject to OpenAI approval and AWS readiness.

  Usage details

  Who can use it: Amazon Bedrock on AWS GovCloud (US-East) for approved government security teams and other eligible customers; Daybreak Red access approval is required.

  How to use it: Complete the Daybreak application or discuss eligibility with your OpenAI/AWS account team. After approval, work with your AWS account team to enable Bedrock access; Daybreak Red approval alone does not grant model access.

  Tip: Have an authorized security expert review the model’s findings and recommended fixes before putting them into production.

  ✓ Verified by official sourceModels: GPT-5.6 CyberSource: Daybreak Red'in AWS GovCloud duyurusu

New in this edition · 13

• 04 Nov
  RetiredMarketing API

  Catalog API /batch endpoint removed in v27

  The Catalog Batch API endpoint is deprecated in v27.0+ and no longer works. Integrations need to migrate to the /items_batch endpoint; /batch will continue to work in v26.0 and earlier until support ends.

  Usage details

  Who can use it: Catalog API integrations, advertiser businesses, and agencies using Marketing API v27.0+.

  How to use it: Replace POST /{catalog-id}/batch calls with POST /{catalog-id}/items_batch and test according to the migration guide.

  Tip: Before moving your API version to v27, test and compare the requests and error responses of both endpoints.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  RetiredConductor Cloud

  New sign-ups and downloads for Conductor have stopped; Cloud shuts down on November 20

  According to Conductor’s official announcement, new sign-ups and app downloads stopped on November 4 following its move to Meta; existing free users cannot upgrade to Pro, while existing Pro users can use the features for free. Conductor Cloud will become read-only on November 13, shut down on November 20, and begin deleting cloud data. Since the source is a Conductor announcement with TechCrunch confirmation, rather than an official Meta channel, verified=false.

  Usage details

  Who can use it: Conductor users; the local Mac app will continue to work, while Cloud users’ migration timeline depends on the November 13–20, 2026 schedule.

  How to use it: Right-click the Cloud workspace and select Transfer to local; use Settings → Account → Export data to export all data, and do so before November 20.

  Tip: Move your cloud workspace and chat archive to a local copy right away; waiting until the read-only phase may limit your export options.

  ! UnverifiedSource: Conductor resmi duyurusu: Meta’ya katılıyoruzSource: TechCrunch: Meta, AI kodlama girişimi Conductor’ı satın aldı
• 04 Nov
  APIMarketing API

  country and location_spec rejected when creating Lookalike audiences

  In Marketing API v27.0+ custom audience creation requests, lookalike_spec.country and lookalike_spec.location_spec are no longer accepted. Lookalike audiences get their delivery locations from the ad sets targeting them; existing audiences and campaigns are unchanged.

  Usage details

  Who can use it: Marketing API v27.0+; creating new Lookalike audiences with POST /act_{ad-account-id}/customaudiences.

  How to use it: Remove the country and location_spec fields from lookalike_spec in new custom audience requests; define the location in ad set targeting.

  Tip: Validate your new audience creation code against the v27 endpoint first, without changing existing audiences.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  RetiredMarketing API

  meta_moment_maker_spec being retired in third-party apps

  In Marketing API v27.0+, meta_moment_maker_spec is no longer returned in GET responses for third-party apps; adding the field to create or update requests results in a version error. Use product_type for Ads Burst and Ads Super Burst inventory; saturation mode is set by the server.

  Usage details

  Who can use it: Marketing API v27.0+; ad set and reach frequency prediction integrations.

  How to use it: Remove meta_moment_maker_spec from read and write flows; use product_type=ADS_BURST or ADS_SUPER_BURST in reach/frequency prediction requests.

  Tip: Validate the new product_type flow in a separate test app before the older API version expires.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 Changelog
• 04 Nov
  RetiredPage Insights

  The Page Insights total_video_consumption_rate metric is no longer returned

  According to Meta’s November 4 v27 announcement, the total_video_consumption_rate metric has not been returned by any Graph API version since early November 2026. No direct replacement for the metric was provided.

  Usage details

  Who can use it: All Graph API versions; GET /{video-id}/video_insights.

  How to use it: Remove the total_video_consumption_rate field from video_insights requests and handle its absence without treating it as an error.

  Tip: Instead of treating an empty or missing metric as zero, show an ‘unavailable’ status in reports.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  SettingMarketing API

  Device expansion enabled by default for eligible Sales ad sets

  Starting with v27.0, for eligible new Sales ad sets targeting a specific device or operating system, Reach more people across other mobile devices is enabled by default if the device field is omitted, and is saved as device=1. Existing ad sets are unchanged; Mobile App conversion location and Special Ad Categories are excluded.

  Usage details

  Who can use it: Marketing API v27.0+; eligible new Sales ad sets targeting a device or OS that don’t use Mobile App conversion location.

  How to use it: To deliver only on the devices and operating systems you selected, send targeting.targeting_automation.individual_setting.device=0.

  Tip: To prevent delivery from silently expanding, explicitly set the device value to 0 or 1 in new ad sets.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  RetiredGraph API

  Six legacy features blocked in new Graph API apps

  Starting November 4, new apps can’t use multi-ID requests, date/time formatting, HTTP method override, field aliases, JSONP callbacks, or asynchronous batch requests; existing apps also can’t add new use cases for them. Existing use in older versions is supported for now; the remaining features will be removed in Graph API v28.0.

  Usage details

  Who can use it: All Graph API versions; new apps and new use cases in existing apps.

  How to use it: Scan your code for uses of ids, date_format, method override, .as(...), callback, and asyncbatch; migrate to supported standard request flows before upgrading to v28.

  Tip: Search your codebase for these six parameters before App Review or a version upgrade, and prevent new usage from being added.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  AvailabilityMeta Ads

  Political and social issue ads can run again in the US

  According to two ad industry sources citing Meta’s official Government and Nonprofits announcement, new US ads about social issues, elections, and politics, as well as edits to existing ads, were re-enabled at 00:00 PT on November 4, 2026. The change is marked as unverified because Meta’s official page could not be accessed as text in this review.

  Usage details

  Who can use it: Accounts advertising about social issues, elections, or politics in the US; after 00:00 PT on November 4, 2026.

  How to use it: Create a new ad or edit an existing one in Ads Manager after that time; separately check authorization, the Paid for by disclaimer, and current Advertising Standards requirements.

  Tip: Reopening does not mean all ads will be automatically approved; verify review and disclaimer status before publishing.

  ! UnverifiedSource: Creative Shift: Meta’nın 2026 siyasi reklam kısıtlamasıSource: Advocacy Aces: Meta’nın 2026 siyasi reklam kısıtlaması
• 04 Nov
  APIMarketing API

  Phone OTP options for lead forms added in Marketing API v27

  With Graph API v27.0, lead forms can use the phone_otp_verification_type field: NOT_ENABLED, OTP_REQUIRED, or OTP_OPTIONAL. OTP_OPTIONAL is available only for eligible ad accounts, and an authorized ad_account_id must be provided when creating a form; the legacy is_phone_sms_verify_enabled field also remains supported.

  Usage details

  Who can use it: Graph API v27.0+; lead form reading and creation flows; OTP_OPTIONAL is available only for eligible ad accounts.

  How to use it: Read the value with GET /{lead-gen-form-id}?fields=phone_otp_verification_type; send phone_otp_verification_type in a POST /{page-id}/leadgen_forms request, and include an eligible ad_account_id for OTP_OPTIONAL.

  Tip: When reading older forms, handle null values and the legacy field together in your client.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  FeatureMarketing API

  Product showcase can be enabled automatically in Marketing API v27

  If the Product showcase setting is absent from a creative, Meta can automatically enable the product_extensions feature when it finds a high-confidence match with a specific product set in the catalog. Because the setting remains on the creative, it also affects subsequent ads that use the same creative.

  Usage details

  Who can use it: Marketing API v27.0+; requests to create new ads and ad creatives with high-confidence catalog matches.

  How to use it: If you don’t want automatic enrollment, send degrees_of_freedom_spec.creative_features_spec.product_extensions.enroll_status=OPT_OUT; if you prefer, check the recommendation before creation using the catalog_suggestions_preview endpoint.

  Tip: If you want full control over brand or product selection, review the preview result and explicitly set OPT_OUT.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  CapabilityGraph API

  AI-generated disclosure management added to Page post editing

  With Graph API v27.0, apps can add or remove AI-generated self-disclosure when editing existing Page posts. Setting is_ai_generated=true adds the label, while false removes it; if the field is omitted, the current value is retained.

  Usage details

  Who can use it: Graph API v27.0+; POST /{post-id} and POST /{page-post-id}; no changes in older versions.

  How to use it: Send the boolean is_ai_generated field in the post edit request; true adds the label, false removes it, and omitting the field retains the current value.

  Tip: If you don’t need to change the label, omit the field to preserve the existing disclosure value.

  ✓ Verified by official sourceSource: Graph API ve Marketing API v27.0 ChangelogSource: Graph API v27.0 ve Marketing API v27.0 duyurusu
• 04 Nov
  FeatureWhatsApp

  Rollout begins for separate, nameable chats for Meta AI in WhatsApp

  WABetaInfo and SammyFans report that in WhatsApp Android beta 2.26.44.6, Meta AI threads are being turned into separate conversations in the chat list. Users can give a new chat a name and custom profile picture, and find these chats using the AI filter in the Chats tab; there is no public official confirmation from Meta or general release date.

  Usage details

  Who can use it: WhatsApp beta for Android 2.26.44.6; some beta testers, phased rollout.

  How to use it: Open a Meta AI conversation, tap the new chat option in the three-dot menu, and set a name and profile picture; open the conversation using the AI filter in the Chats tab.

  Tip: Use a separate chat for each topic to reduce context mixing and the need to search for old threads.

  ! UnverifiedSource: WABetaInfo: WhatsApp Android’de Meta AI sohbetlerini rollout ediyorSource: SammyFans: WhatsApp kullanıcıları Meta AI ile ayrı sohbetler oluşturabiliyor
• 03 Nov
  SettingPerformance Analyzer

  Viewing live metrics no longer starts recording automatically

  In Meta VR Developer Hub 6.6.0, viewing live metrics no longer starts recording on its own. This prevents developers who only want real-time monitoring from generating unnecessary traces.

  Usage details

  Who can use it: Meta VR Developer Hub for Mac 6.6.0 and Performance Analyzer users.

  How to use it: Open the live metrics view in Performance Analyzer; start recording separately when a trace is needed.

  Tip: Use only live metrics to reduce storage consumption during short diagnostic sessions.

  ✓ Verified by official sourceSource: Meta VR Developer Hub 6.6.0 sürüm notları

New in this edition · 33

• 04 Nov
  AgentKimi Code Desktop

  Desktop subagents can control the built-in browser with permission

  Running subagents can now use Desktop’s built-in browser with user permission. This lets parallel tasks handle page navigation and verification separately from the main Agent flow.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0; subagent tasks that require user approval and browser access.

  How to use it: Update Desktop to v1.3.0, assign the subagent a task that includes browser verification, and approve access in the permission card that appears.

  Tip: Use Always Ask for initial trials, and check which pages the subagent accesses before the task is complete.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code Desktop görev yürütme kılavuzu
• 04 Nov
  SettingKimi Code CLI

  Change the model and thinking level while Kimi Code is running

  The `/model` and `/effort` commands can now be used while the Agent is running. The model call already in progress keeps its existing settings; the selected value applies to the next turn, and Compact and continue waits for the current turn to finish.

  Usage details

  Who can use it: Kimi Code CLI v2.5.0 and later.

  How to use it: During a running turn, enter `/model` or `/effort` and select a new model or the `low`, `high`, or `max` level.

  Tip: When changing the model midway through a long turn, keep in mind that it affects the next turn, not the current result.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Slash Commands referansı
• 04 Nov
  CapabilityKimi Code CLI

  Kimi Code CLI runs remote environments over SSH, Docker, and custom commands

  Agent can now run file and Shell operations on an SSH machine, in a Docker-compatible container, or through a custom launcher command; the CLI, model requests, credentials, and session state remain local. Windows remote targets are also supported over SSH and custom commands.

  Usage details

  Who can use it: Kimi Code CLI v2.5.0; SSH, Docker, or custom-command environments; Docker is not supported on Windows targets.

  How to use it: Define `[environments.dev-box]` in `~/.kimi-code/config.toml`, then run `kimi --environment dev-box` or `kimi --environment dev-box --environment-cwd /path` for a different directory.

  Tip: Install the Kimi Code executor on the remote machine in advance; commands don’t silently fall back to the local machine if the connection drops.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Remote environments kılavuzu
• 04 Nov
  FeatureKimi Code Desktop

  Kimi Code Desktop lets you add multiple working directories to projects

  With Desktop v1.3.0, you can select a primary directory and additional source folders when creating a project. The directory list can be edited later, and older additional directory configurations can be imported.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0; local multi-directory projects.

  How to use it: When creating a new project, select a primary directory and add source folders; later, edit the project directory list or import a legacy additional directory configuration.

  Tip: Choose the most stable root as the primary directory for shared configuration and conflict resolution.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code güncel değişiklik özeti
• 04 Nov
  Plugin / IntegrationKimi Code CLI

  Kimi Code’s plugin marketplace and built-in MCP OAuth flow updated

  Kimi Code 2.5.0 supports plugin marketplace interactions and built-in OAuth authorization for MCP servers within plugins. Official, Curated, and custom URL sources can all be used in the same manager.

  Usage details

  Who can use it: Kimi Code CLI v2.5.0 and Desktop plugin settings; plugins are user-scoped.

  How to use it: Open `/plugins` in the TUI and install from the Official or Curated tab; use `/plugins marketplace <source>` for a custom source, and start the `/mcp-config` flow for MCP OAuth.

  Tip: After installing or enabling a plugin, use `/reload` or `/new` to apply the change to a new session.

  ✓ Verified by official sourceSource: Kimi Code 2.5.0 resmi GitHub sürümüSource: Kimi Code 2.5.0 canary resmi sürümüSource: Kimi Code Plugins kılavuzu
• 04 Nov
  FeatureKimi Code

  Kimi Code can fork from a previous turn or a running session

  A session can now be copied independently from a selected previous turn; a fork can also be created while the source session is running. This makes it possible to try alternative implementation approaches without disrupting the current flow.

  Usage details

  Who can use it: Kimi Code CLI v2.5.0 and Kimi Code Desktop v1.3.0.

  How to use it: Use `/fork` in the CLI; in Desktop, select `Fork from here` from the session or previous-turn menu.

  Tip: Make major experimental changes in a fork to keep the main session’s context and goal clear.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi GitHub sürümüSource: Kimi Code Desktop 1.3.0 changelog’uSource: Kimi Code oturum kılavuzu
• 04 Nov
  RetiredKimi Code Server API

  Kimi Code REST task and status endpoints removed

  `/sessions/{id}/tasks`, status and snapshot endpoints, along with the legacy WebSocket subscribe/unsubscribe flow, have been removed. Task output and control have moved to WebSocket `task.watch`, `task.unwatch`, `task.stop` and `task.background` requests.

  Usage details

  Who can use it: Third-party clients using Kimi Code’s local REST/WebSocket API; after 2.5.0.

  How to use it: Remove legacy REST task calls and monitor tasks with `task.watch` and `task.unwatch` over `/api/v1/ws`; send `task.stop` and `task.background` to stop a task or move it to the background.

  Tip: In migration code, check the server version and capabilities via `/api/v1/meta`, and show users a clear error message for legacy endpoints.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Server API referansı
• 04 Nov
  APIKimi Code Server API

  Kimi Code WebSocket stream carries Agent and task status

  The local server WebSocket now sends live status, settings, context usage and task changes through `agent`, `usage`, `goal`, `todo`, `task.list` and `task` frames. The subscription flow has been updated to use `subscribe_v2` and resubscribe after reconnecting.

  Usage details

  Who can use it: Integration developers using the Kimi Code CLI 2.5.0 local REST/WebSocket server.

  How to use it: Start `kimi web`; read the live `/openapi.json` and `/asyncapi.json` schemas for your version and use `subscribe_v2` over `/api/v1/ws`.

  Tip: Don’t bind your client to a fixed list of fields; reread the live OpenAPI and AsyncAPI schemas with every Kimi Code version.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Server API referansı
• 04 Nov
  AvailabilityKimi K3.1 Pro

  Kimi K3.1 Pro open weights released on Hugging Face

  Moonshot AI has released Kimi K3.1 Pro’s full model weights, making it possible to run the model on your own infrastructure. The official model card says it uses 2.8 trillion total parameters, 104 billion active parameters, and a 1,048,576-token context window; it recommends vLLM, SGLang, and TokenSpeed.

  Usage details

  Who can use it: Available to developers and researchers via Hugging Face; the Kimi K3.1 Pro License terms apply.

  How to use it: Download the `moonshotai/Kimi-K3.1-Pro` repository from the Hugging Face model card and serve it using one of the vLLM, SGLang, or TokenSpeed examples in the card.

  Tip: Before putting a 1-million-token context window into production, use a short pilot to measure GPU, latency, and memory costs on your own setup.

  ✓ Verified by official sourceModels: Kimi K3.1 ProSource: Kimi.ai’nin ağırlık yayın duyurusuSource: Kimi K3.1 Pro resmi Hugging Face model kartı
• 04 Nov
  CapabilityKimi Code

  Background tasks show live notifications and detailed output

  The CLI displays a notification while a background task is producing a result; Desktop brings running tasks to the foreground and accumulates detailed output live. Agent and subagent status, model, and result information can be tracked in the task view.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and Desktop v1.3.0.

  How to use it: Open `/tasks` in the CLI; in Desktop, select the running task from the background work bar above Composer or the Agent/Subagent panel.

  Tip: After moving a long task to the background, check its live status and output first rather than starting the same task again.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Desktop 1.3.0 changelog’uSource: Kimi Code görev yürütme kılavuzu
• 04 Nov
  CapabilityKimi Code Desktop

  Desktop displays readable computer-use and tool-selection cards

  Computer-use and tool-selection calls are now shown with clear titles, parameters, and result summaries. Users can see what each tool did without having to parse raw technical output.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: When the Agent makes a computer-use or tool-selection call, open the card in the conversation and review its title, parameters, and result summary.

  Tip: Before approving, check the target application, URL, and parameters to be entered in the card.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’u
• 04 Nov
  FeatureKimi Code Desktop

  Desktop adds diagnostics export and a startup recovery screen

  Desktop now allows you to export diagnostics by packaging logs and crash reports in a single ZIP. When startup is slow or fails, the loading page also offers options to export diagnostics and restart.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: Select `Help → Export Diagnostics…`; if the app won’t open, use the Export Diagnostics option on the loading screen.

  Tip: Before sending the ZIP to the support team, check whether it contains source code, command output, or file paths.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code Desktop güncel değişiklik özeti
• 04 Nov
  FeatureKimi Code Desktop

  A new right-hand panel for session information is available in Desktop

  The new right-hand panel brings together supporting information for the current session in one place. The panel stays in tabs alongside the session, and users can open or close details as needed.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: Open the right-hand panel on the session screen and select the Session Information tab; expand the details if they’re collapsed by default.

  Tip: In multi-directory projects, check the session context in the panel before taking action.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code Desktop arayüz kılavuzu
• 04 Nov
  APIKimi Code CLI

  Kimi Code displays backend model errors verbatim

  The CLI now displays backend model error messages directly instead of reinterpreting them. Quota status is no longer inferred from the error text; explicit retryability information is required for retry policies.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and custom provider/model integrations.

  How to use it: When a model call fails, handle the original backend message and retry status directly from the CLI output.

  Tip: Apply automatic retries only to errors explicitly marked as retryable.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Server API referansı
• 04 Nov
  RetiredKimi Code Desktop

  Kimi Code Desktop now requires macOS 13 or later

  With Desktop v1.3.0, macOS 13 is now the minimum supported version; macOS 12 can no longer receive new updates. Windows and x86_64 Linux packages continue to be available.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0; macOS 13 and later, Windows, x86_64 Linux.

  How to use it: Check your macOS version in About This Mac; if you’re on macOS 12, upgrade the operating system or stay on a compatible older version before installing v1.3.0.

  Tip: For enterprise Mac deployments, plan the operating system upgrade before updating Desktop.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code Desktop hızlı başlangıç kılavuzu
• 04 Nov
  Plugin / IntegrationKimi Code for VS Code

  Kimi Code for VS Code automatically refreshes the model catalog

  The VS Code integration has been updated to automatically refresh the provider model catalog. This reduces the need to manually reimport catalog changes when a provider adds a new model.

  Usage details

  Who can use it: Kimi Code for VS Code users with the Kimi Code 2.5.0 canary 2026-11-04 change.

  How to use it: Update Kimi Code for VS Code and reload the VS Code window; the list refreshes automatically when the provider or model catalog is opened.

  Tip: After automatic refresh, check the model’s context and capability information before using it in a production request.

  ✓ Verified by official sourceSource: Kimi Code 2.5.0 canary resmi GitHub sürümüSource: Kimi Code for VS Code hızlı başlangıç kılavuzu
• 04 Nov
  SettingKimi Code CLI

  Kimi Code disables global full-text search by default

  Kimi Code no longer creates a global full-text index by default to scan past sessions in the background. Web search is therefore unavailable by default, but can be re-enabled with an explicit setting.

  Usage details

  Who can use it: Kimi Code CLI/Web 2.5.0; users who want to search all past sessions.

  How to use it: Set `[database] global_search = true` in `config.toml` or provide the `KIMI_CODE_GLOBAL_SEARCH=1` variable.

  Tip: Keep the default disabled behavior for sensitive projects; enable search only for trusted local sessions when needed.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code yapılandırma referansı
• 04 Nov
  AgentKimi Code CLI

  Kimi Code can change the current Goal objective within a session

  Users can edit the objective text of an ongoing Goal in place without starting a new session. This preserves existing progress and session context when the objective’s scope changes.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 canary 2026-11-04 and versions that include this change.

  How to use it: Replace the active objective with a new one using `/goal replace <objective>`; the Goal’s status and progress remain in the same session.

  Tip: When changing the objective, include a measurable completion condition in the new text.

  ✓ Verified by official sourceSource: Kimi Code 2.5.0 canary resmi GitHub sürümüSource: Kimi Code slash komutları referansı
• 04 Nov
  SettingKimi Code CLI

  Kimi Code subagent model selection and `secondary_model.force` behavior changed

  `secondary_model.force` is now ignored and triggers a warning; implicit `primary` selection has been removed for single-model pools. Model selection for Agent/AgentSwarm spawns is available again when two or more aliases are defined.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0; users of custom subagent model pools.

  How to use it: List at least two model aliases under `[secondary_model.models]` in `config.toml`; use only `default_model` for a fixed model and do not add `force`.

  Tip: Define a fast model and a powerful model with descriptive aliases to distinguish the cost and speed of subtasks.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code yapılandırma referansı
• 04 Nov
  SettingKimi Code Desktop

  Queued Desktop messages use the session’s current permission mode

  Messages queued while the Agent is running now use the session’s permission mode when they’re processed, rather than the setting that was in effect when they entered the queue. This keeps behavior consistent if the user changes the permission mode and runs a pending message later.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: Send a message while the Agent is running, then change the permission mode below Composer and wait for the message to be processed from the queue.

  Tip: When side-effecting operations are queued, explicitly check Always Ask before choosing Ask When Needed for automated operations.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code Desktop görev ve izin kılavuzu
• 04 Nov
  Plugin / IntegrationKimi Code CLI

  The tool catalog refreshes in the same turn when MCP permissions or tools change

  After MCP OAuth is completed or the tool configuration changes, Kimi Code refreshes the available tools within the current turn. There’s no need to wait for a new session or manually restart.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and sessions using an MCP server.

  How to use it: Authorize or modify the MCP server with `/mcp-config`; expect Agent to use the new tool catalog in the same turn.

  Tip: After authorization, check that Agent has refreshed its tool list before starting a new turn.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Plugins kılavuzu
• 04 Nov
  APIKimi Code CLI

  Invalid `max_tokens` error fixed for custom model registries

  The output token limit declared by models retrieved from a custom registry is now saved. This fixes the 400 `Invalid max_tokens` error seen with some models from custom registries.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and models imported from a custom registry.

  How to use it: Re-import the registry with `kimi provider add <url>` or refresh the provider model list; requests will use the saved output limit.

  Tip: Match the model output limits in the registry’s `api.json` to the actual service limits.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi GitHub sürümüSource: Kimi Code komut referansı
• 04 Nov
  CapabilityKimi Code CLI

  Remote file reading, writing, searching, and completion are faster

  Network round trips have been reduced for file reading, editing, writing, searching, and completion in remote environment sessions. This change reduces interaction latency, especially in SSH- and container-based projects.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 remote environment sessions.

  How to use it: Define a remote environment and start a session with `kimi --environment <id>`; no additional settings are required.

  Tip: Referencing a relevant folder as a whole instead of many small files saves even more network calls.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi GitHub sürümüSource: Remote environments kılavuzu
• 04 Nov
  SettingKimi Code CLI

  Option to bypass the workspace trust prompt with `require_trust = false` added

  Kimi Code CLI now offers a configuration option to bypass workspace trust prompts. When `[workspace] require_trust = false` is set, new workspaces don’t require trust approval.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0; trusted automation environments.

  How to use it: Add `[workspace]` and `require_trust = false` to `~/.kimi-code/config.toml`, then start a new session.

  Tip: Enable this setting only for folders whose contents and execution permissions are controlled.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code yapılandırma referansı
• 04 Nov
  CapabilityKimi Code Desktop

  Desktop File Tree can expand symbolic link directories

  File Tree can now open directories represented by symbolic links and browse the files inside them. This broadens file discovery in projects that use monorepos and shared source folders.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: Open the File Tree panel and expand a directory marked with the symbolic link icon.

  Tip: Open only the links you need to avoid unnecessary scanning of circular symbolic links.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’u
• 04 Nov
  FeatureKimi Code Desktop

  Desktop opens Markdown files in the default system app

  Kimi Code Desktop v1.3.0 can open Markdown documents in the default system app as well as in its in-app preview. This makes it easier to quickly review a document in an external editor or Markdown reader.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0; macOS, Windows, and Linux desktop users.

  How to use it: Open a Markdown file in File Preview and select `Open` from the file actions.

  Tip: Use Open and Preview in sequence to compare the source Markdown with the rendered preview.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’uSource: Kimi Code görev yürütme kılavuzu
• 04 Nov
  SettingKimi Code Desktop

  Desktop adds a Follow default model option to subagent model settings

  Subagent model settings now include the `Follow default model` option, which follows the default model in the main configuration. This means you don’t need to manually sync the subtask setting when the main model changes.

  Usage details

  Who can use it: Kimi Code Desktop v1.3.0.

  How to use it: Open the subagent model settings in Settings and select `Follow default model`.

  Tip: Use a fixed model for fast, cost-controlled subtasks and Follow default model for tasks that vary.

  ✓ Verified by official sourceSource: Kimi Code Desktop 1.3.0 resmi changelog’u
• 04 Nov
  OtherKimi Code CLI

  Invalid additional workspace paths are now shown in error messages

  When starting the TUI, an invalid additional workspace directory no longer causes a silent exit; the problematic path is shown to the user. This makes it easier to identify which folder caused a configuration error in multi-directory projects.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and users of multi-directory workspaces.

  How to use it: Start the CLI; if an error occurs, correct the displayed directory path and try again with `config.toml` or the session’s additional directory list.

  Tip: Specify additional directories using absolute paths that exist; remove outdated paths after moving directories.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi GitHub sürümü
• 04 Nov
  CapabilityKimi Code CLI

  Kimi Code removes Edit and `sed` restrictions for incremental edits

  The rules requiring the Edit tool for incremental file edits and prohibiting Bash `sed` have been removed. The Agent can use different editing methods when appropriate.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and later.

  How to use it: No additional settings are required; assign a file-editing task and review the Edit or Bash approach chosen by the Agent in the approval flow.

  Tip: Even though the restriction has been removed, review the diff for large changes and prefer small, reversible steps.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’u
• 04 Nov
  CapabilityKimi Code CLI

  Kimi Code shows todo update reminders less often

  The Agent now receives a reminder if it hasn’t updated the todo list for 10 steps instead of 5, and avoids unnecessary warnings when the list still matches progress. This reduces repeated planning messages on long tasks.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0.

  How to use it: No additional setup is needed; start a long task and track updates in `/tasks` or the todo card.

  Tip: For long Goal tasks, manually updating the todo list at key milestones is more reliable than waiting for an automatic reminder.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’u
• 04 Nov
  Plugin / IntegrationKimi Code CLI

  Canceling MCP OAuth no longer interrupts other authorization attempts

  Canceling an MCP OAuth authorization no longer accidentally stops another authorization attempt that is pending at the same time. Authorization flows work more independently for teams setting up multiple MCP servers.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0 and multi-server setups using MCP OAuth.

  How to use it: Authorize multiple servers through `/mcp-config`; when you cancel one flow, check that the other continues.

  Tip: Use separate authorization windows and clear names for each server to avoid confusing which OAuth flow you’re canceling.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code Plugins kılavuzu
• 04 Nov
  OtherKimi Code CLI

  Steer messages appear dimmed until the Agent reads them

  Steer messages sent to a running Agent now appear dimmed until the Agent reads them, then return to normal. The status carries over to subsequent turns, making it easier to tell whether a message was actually processed.

  Usage details

  Who can use it: Kimi Code CLI 2.5.0.

  How to use it: Send a message with `Ctrl-S` while the Agent is running, and watch the dimmed/read status on the message card.

  Tip: For critical directions, don’t pile on new requests until you confirm the Agent has applied the message in its next response.

  ✓ Verified by official sourceSource: Kimi Code CLI 2.5.0 resmi changelog’uSource: Kimi Code etkileşim kılavuzu
• 04 Nov
  CapabilityKimi Code CLI

  The local FetchURL tool can access private network addresses

  Kimi Code Agent’s local FetchURL tool has been expanded to access private addresses. This makes it possible to test local development services or internal endpoints through Agent’s own FetchURL flow.

  Usage details

  Who can use it: Versions of Kimi Code CLI 2.5.0 that include the 2026-11-04 canary change; network access depends on the machine’s permissions.

  How to use it: Ask Agent to inspect an accessible local or private network URL; the FetchURL tool is used if network and security policies allow it.

  Tip: Limit private network access to trusted URLs; don’t expose credentials or administrative endpoints to Agent.

  ✓ Verified by official sourceSource: Kimi Code 2.5.0 canary resmi GitHub sürümü

New in this edition · 2

• 05 Nov
  AvailabilityGrok Bot

  Grok Bot phone calls are limited to eligible US accounts and US numbers

  An official post dated November 5 confirms that Grok Bot can make real phone calls, handle menus, hold times, and transfers, and provide live transcripts. Users can respond when needed and receive a summary at the end of the call; the rollout is still ongoing and is currently limited to eligible US accounts calling US numbers.

  Usage details

  Who can use it: Eligible US accounts; US numbers only; gradual rollout. The post does not confirm access for accounts or numbers outside the US.

  How to use it: Update the Grok Bot app; tell Bot whom to call and what you want it to do; respond in the chat box if it asks for information during the call.

  Tip: Before the call, clarify the name, account details, and desired outcome for Bot; handle final approvals for charges, cancellations, or contracts yourself.

  ✓ Verified by official sourceSource: Grok’un 5 Kasım 2026 tarihli resmi gönderisi
• 03 Nov
  APISpaceXAI TypeScript SDK

  Watermarked fields added in SpaceXAI TypeScript SDK v0.2.8

  The v0.2.8 release dated November 3 added response.watermarked to Responses API responses and stream.watermarked to streams. The fields return a boolean based on the x-watermark: true header; when retryBeforeOutput is used, the header from the attempt that produced the result is used.

  Usage details

  Who can use it: TypeScript developers working with the Responses API using SpaceXAI TypeScript SDK v0.2.8; the SDK requires Node.js 22.13 or later and an ESM project.

  How to use it: Upgrade with npm install @xai-official/sdk@0.2.8; read response.watermarked from the responses.create() result and stream.watermarked from a stream with stream: true. Enabling or disabling Watermarking is a team-level Console setting, not an SDK parameter.

  Tip: Write the boolean field directly to the audit log; do not try to parse the x-watermark header separately or add this state to the request body.

  ✓ Verified by official sourceSource: SpaceXAI TypeScript SDK v0.2.8 sürüm notlarıSource: Watermarked alanlarının SDK commit ayrıntıları

New in this edition · 1

• 04 Nov
  ModelTencent Hy

  Tencent Hy4 announced: visual understanding and computer use for the first time

  Tencent announced its next-generation large language model, Hy4. For the first time, the model adds visual understanding (fine-grained image processing, complex chart analysis, mathematical visual reasoning) and computer use through the command line and graphical interfaces; it also improves instruction following, long context, and long-running agent tasks. It ranks 14th in the unique rankings with a score of 35 on the Artificial Analysis Intelligence Index v5.0.1 (1M-token context, text+image input).

  Usage details

  Who can use it: For now, it is available through Tencent WorkBuddy and WeCom integration; according to Artificial Analysis, there is no public API. Tencent says it will release the weights and system cards as open source “in the coming days.”

  How to use it: WorkBuddy users can use Hy4 directly for office, visual design, 3D scene, and web development tasks; when connected to WeCom, it tracks projects using chat, email, document, calendar, and meeting data.

  Tip: When the weights are published on Hugging Face (huggingface.co/tencent), check the license terms before trying them on your own infrastructure; for now, treat it as a closed model.

  ✓ Verified by official sourceModels: Tencent Hy4Source: Tencent Launches Hy4 (PR Newswire)Source: Tencent Hy4 – Artificial Analysis

New in this edition · 35

• 04 Nov
  AvailabilityQwen Code

  Local Docker recipe added for /verify-pr

  Qwen Code published a local `/verify-pr` Docker recipe that runs untrusted PR code in a container. Model credentials stay on the host, behind a chat-completion proxy.

  Usage details

  Who can use it: Developers using a local Qwen Code installation, especially those who want to isolate untrusted PRs.

  How to use it: Use the local `/verify-pr` Docker recipe in the Qwen Code repository; keep PR code in the container and model credentials on the host proxy.

  Tip: Limit the container running PR verification to only the necessary file and network permissions.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  Plugin / IntegrationQwen Code Channels

  Linked PR events arrive in the session as untrusted prompts

  GitHub events for linked PRs are now passed to the relevant session as untrusted prompts. Responses remain private until explicitly published.

  Usage details

  Who can use it: Users who link GitHub PRs to Qwen Code Channel sessions.

  How to use it: Link a PR to a Channel session; review incoming GitHub events as untrusted prompts and publish the response only when needed.

  Tip: Treat automatically received PR content as external data to verify, not as instructions.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  FeatureQwen Code Web Shell

  Filterable live timeline added for Channel events

  A separate page showing Channel events has been added to Web Shell. It presents events, skip reasons, replies, task transitions, and polling results in a filterable, automatically refreshed timeline.

  Usage details

  Who can use it: Teams using Qwen Code Web Shell Channel.

  How to use it: Open the Channel events page in Web Shell; use the filters and follow the timeline with live refresh enabled.

  Tip: When investigating the source of a channel issue, narrow the skip reason and task transition filters first.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  APIQwen Code Managed Agent

  Managed Agent supports full model configuration at session startup

  Managed Agent now accepts the full model configuration when a session is created. This allows you to specify not only the model name but also other model options the session needs in the initial request.

  Usage details

  Who can use it: Developers using the Qwen Code Managed Agent session creation flow.

  How to use it: Send the full model configuration in a new Managed Agent session creation request; populate the model fields in the session creation schema for the version you’re using.

  Tip: Setting model options at session startup helps ensure consistent behavior in reconnection and supervised generation flows.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  SettingQwen Code

  Project hooks require explicit folder or workspace trust

  Qwen Code now requires explicit folder or workspace trust before running repository-sourced hooks, permission grants, and project language servers. If trust isn’t granted, a warning appears in the interactive session.

  Usage details

  Who can use it: All interactive Qwen Code sessions that use `.qwen/settings.json` or project hooks in a repository.

  How to use it: Review the trust prompt when you open a project; grant folder/workspace trust only to a folder whose source you’ve verified, then run the hooks.

  Tip: Before granting trust in unknown repositories, inspect `.qwen/settings.json` and the language server configuration.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  SettingQwen Code

  Remote connection settings for qwen serve without a restart

  Qwen Code added the `Settings → Connections → Allow remote connections` setting. It lets you manage remote access, token usage, and client disconnections without restarting the service.

  Usage details

  Who can use it: Qwen Code installations using `qwen serve`.

  How to use it: Enable `Settings → Connections → Allow remote connections`; configure a token for remote client access and manage the connection without restarting.

  Tip: Enable remote access only for trusted networks and clients; do not put the token in source code or a shared configuration file.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  APIQwen Code Anthropic-compatible API

  Anthropic-compatible requests preserve configured timeouts

  Anthropic-compatible requests now honor configured transport timeout and dispatcher values. This prevents requests from being cut off early when a longer API timeout is set.

  Usage details

  Who can use it: Developers using Anthropic-compatible provider endpoints in Qwen Code.

  How to use it: Set the desired transport timeout in the Anthropic-compatible provider configuration; verify that the value is preserved with a long-running request.

  Tip: When increasing the timeout, also monitor the provider’s actual server limit and the total task budget separately.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  Plugin / IntegrationQwen Code Channels

  Channel polling cursors are now scoped per workspace

  GitHub, GitLab, and DWS polling cursors are now stored separately for each workspace. Existing positions are preserved during upgrades, and channel state is stored with stricter permissions.

  Usage details

  Who can use it: Teams using GitHub, GitLab, or DWS Channel across multiple workspaces.

  How to use it: Upgrade Qwen Code to v0.26.7; verify Channels separately in each workspace and check that new events are not fetched again after the upgrade.

  Tip: Compare the latest event time in each workspace after upgrading; reprocessing old events may indicate a cursor scoping issue.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code

  Tool selection errors fixed for four Claude-compatible models

  Qwen Code fixed tool selection errors with Claude Opus 5.5, Claude Sonnet 5.5, Fable 5.1, and Mythos 5.1. Unsupported forced tool selection no longer causes failed calls with these models.

  Usage details

  Who can use it: Users running these models in tool-using workflows through Qwen Code.

  How to use it: Select the relevant model and run a prompt that uses a tool; verify that the call completes with non-forced tool choice behavior.

  Tip: When the model provider is updated, recheck the same tool call with a short regression prompt.

  ✓ Verified by official sourceModels: Claude Opus 5.5, Claude Sonnet 5.5, Fable 5.1, Mythos 5.1Source: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  Plugin / IntegrationQwen Code Generated UI

  Node-based Generated UI MCP runs in the Electron IDE companion

  Qwen Code’s built-in Generated UI MCP server now runs properly in Node mode when launched from a CLI Electron-based IDE companion. This makes the Generated UI integration available through this client path.

  Usage details

  Who can use it: Generated UI users running Qwen Code with an Electron-based IDE companion.

  How to use it: Launch Qwen Code CLI from the Electron IDE companion and test a tool call that produces MCP Apps/Generated UI output.

  Tip: Verify the client’s MCP Apps display support beforehand; a running server alone does not mean visual cards will be displayed.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code Web Shell

  GitHub channel editor now covers repo, mention, and polling settings

  The GitHub channel editor, which was previously made available, has been expanded; repository access, mention behavior, and notification polling settings can now also be edited from Web Shell. This reduces the need to configure these settings manually.

  Usage details

  Who can use it: Web Shell GitHub Channel users; the expanded scope of the previously available GitHub channel settings feature.

  How to use it: In Web Shell, edit the GitHub repository access, mention behavior, and notification polling fields in the Channel editor.

  Tip: Verify mention and polling settings in a small test repository to reduce notification noise.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code

  grep -n options are correctly recognized in permission analysis

  Shell permission analysis now recognizes `grep -n` and `grep --line-number` as options. This prevents grep calls intended to read files from incorrectly bypassing permission checks.

  Usage details

  Who can use it: All local and Managed sessions using Qwen Code shell tool permissions.

  How to use it: Run a shell call using `grep -n` or `grep --line-number` and check that tool approval/permission analysis correctly classifies the file-reading behavior.

  Tip: Rather than broadening shell permissions, review the tool description to confirm that the command is truly read-only.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  FeatureQwen Code Web Shell

  Images can be attached by pasting an HTTP(S) URL

  When a single HTTP(S) image URL is pasted into Web Shell, the image is now attached. Users also receive a warning when a file paste is unavailable.

  Usage details

  Who can use it: Desktop and touch clients using prompts with image attachments in Web Shell.

  How to use it: Copy a single image HTTP(S) URL to the clipboard and paste it into the Web Shell composer; check that the attachment was added before sending.

  Tip: Use a permanent, authorized image URL the model can access instead of a temporary or access-protected URL.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  FeatureQwen Code Web Shell

  Managed Web Shell entry and session error notifications improved

  A Managed host entry has been added to Web Shell. The Managed empty state layout has been improved, and session failure and unconfirmed submission notifications are now grouped above the composer.

  Usage details

  Who can use it: Managed Web Shell users.

  How to use it: Open the Managed host entry on the Web Shell home page; if the session doesn’t load, follow the error or verification notification above the composer.

  Tip: If a session isn’t added, check the unconfirmed submission warning above the composer before trying again.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  SettingQwen Code

  New settings for model catalog refresh and usage telemetry

  Qwen Code adds the operator-wide `model.catalogRefresh` setting. When `privacy.usageStatisticsEnabled:false` is used, background model catalog refreshes stop, but local model metadata remains available.

  Usage details

  Who can use it: Operators using Qwen Code v0.26.7 and installations managing privacy settings.

  How to use it: Set `model.catalogRefresh` in the operator configuration; use `privacy.usageStatisticsEnabled:false` to disable background catalog refreshes.

  Tip: When disabling catalog refreshes, make sure the local catalog data is up to date; otherwise, new model names may not appear.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code

  262.144-token context is preserved for OpenRouter qwen3-coder:exacto

  The `qwen/qwen3-coder:exacto` identifier on OpenRouter now preserves the base model name and is recognized with a 262.144-token context window instead of 200K. This ensures the model’s actual context capacity is correctly passed to Qwen Code.

  Usage details

  Who can use it: Users of OpenRouter `qwen/qwen3-coder:exacto` through Qwen Code.

  How to use it: Select OpenRouter as the model provider and use `qwen/qwen3-coder:exacto` as the model identifier; verify the capacity with a long-context prompt.

  Tip: Before filling the context window, also check the provider’s actual limits and pricing.

  ✓ Verified by official sourceModels: qwen/qwen3-coder:exactoSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code review

  PR review report now shows final cohort and endpoint details

  PR review routing now selects the cohort before the model and reports the final bucket, cohort, model, and endpoint hostname. This makes it clear which route canary or experimental routing took.

  Usage details

  Who can use it: Teams using the Qwen Code PR review workflow.

  How to use it: Run a PR review; check the bucket, cohort, model, and endpoint hostname fields in the completed report.

  Tip: When model quality changes, compare the cohort and endpoint in the report first to distinguish routing differences.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code

  prompt.submit hooks are now available in the Ink terminal

  Mods `prompt.submit` hooks now run in the Ink terminal before the typed prompt is processed. Hooks can see the composer text and expanded pasted content; this extends the previous headless support to the terminal.

  Usage details

  Who can use it: Qwen Code Mods developers using the Ink terminal.

  How to use it: Define a `prompt.submit` hook in a Mod and verify in the Ink terminal that it observes or transforms the text before the prompt is submitted.

  Tip: For hooks that transform text, check the logs to see which form of the text is sent to the model after pasted content is expanded.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code Desktop

  RHEL 10 EPEL bootstrap RPM pinned

  The EPEL bootstrap RPM used in the Qwen Code Desktop installation flow on RHEL 10 has been pinned. This fix improves installation reliability for the RHEL 10 RPM distribution added on November 2.

  Usage details

  Who can use it: Users of RHEL 10 x64 Desktop packages.

  How to use it: Update Qwen Code Desktop to v0.26.7 or later and rerun the RHEL 10 RPM installation.

  Tip: Before changing the EPEL RPM separately in your enterprise mirror, verify that the release uses the pinned package.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code ACP

  unifiedAgentLoop expanded to ACP scheduled and background turns

  When `experimental.unifiedAgentLoop` is enabled, ACP cron turns and background notification turns also use the shared loop. This preserves shared turn, tool grouping, and stop behavior in these turns.

  Usage details

  Who can use it: ACP clients running scheduled tasks or background notifications; scope expansion in v0.26.7.

  How to use it: Enable `experimental.unifiedAgentLoop`; start a cron or background-notification turn through ACP.

  Tip: To prevent duplicate notifications in background turns, compare the results of the old and new loops with the same client.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code ACP

  unifiedAgentLoop now also covers ACP Goal turns

  With the experimental `unifiedAgentLoop` enabled, ACP Goal turns now use the turns, tool groups, and stop points in the shared loop. This expands the previously documented ACP shared-loop coverage to the Goal flow.

  Usage details

  Who can use it: ACP clients using `experimental.unifiedAgentLoop`; Qwen Code v0.26.7.

  How to use it: Enable `experimental.unifiedAgentLoop` in Qwen Code experimental settings and run the ACP Goal flow.

  Tip: Before putting the Goal flow into production, test that tool groups and stop point behavior are compatible with your client.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code qwen serve

  Remote access errors no longer expose internal details to callers

  The remote access error-handling flow no longer displays internal error details to callers. Various remote access reliability and security issues were also fixed in the same release.

  Usage details

  Who can use it: Clients and service owners that use `qwen serve` remote access.

  How to use it: After enabling remote access, review errors on the client side; for detailed diagnostics, check the server logs as an authorized operator.

  Tip: Keep safe errors returned to users and authorized server logs in separate channels.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Workspace file uploads can now run concurrently

  Workspace uploads can now proceed at the same time. The system continues to reject occupied destinations and enforce directory capacity protections.

  Usage details

  Who can use it: Managed Agent users using Qwen Code Workspace.

  How to use it: Submit multiple files at the same time on the Workspace upload screen; rejection behavior for occupied destinations is unchanged.

  Tip: When uploading many files, check destination directory names in advance; conflicting destinations will still be rejected.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  FeatureQwen Code ACP

  Large tool results are shown as condensed previews in ACP session replays

  The live ACP tool result view now shows large file-editing results as saved session previews instead of full old and new content. This reduces unnecessarily long output in session replays.

  Usage details

  Who can use it: Users of ACP session replays or Managed Web Shell history.

  How to use it: After a large file-editing turn, open the ACP session replay view; inspect the saved preview instead of the full result.

  Tip: When you need the full diff, open the saved session or artifact linked from the preview; don’t use the replay screen as a raw diff repository.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  AvailabilityQwen Code Web Shell

  Navigation on touch devices keeps focus in chat

  Web Shell navigation on touch devices now keeps focus in the chat panel instead of automatically opening the composer keyboard. Desktop focus behavior is unchanged.

  Usage details

  Who can use it: Users of Qwen Code Web Shell on touch devices.

  How to use it: Switch between Web Shell sections on a touch device; use the chat panel while it remains focused without the composer keyboard opening unintentionally.

  Tip: If you need the keyboard to open, tap the composer separately; this behavior changes only navigation focus.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Harness admission, recovery, and Hook checks are faster

  Standalone Harness admission, recovery, cleanup, and Hook admission checks no longer use the tenant placement lock in safe cases. Managed Agent performance improves while decision outcomes remain unchanged.

  Usage details

  Who can use it: Teams using Standalone Harness and Managed Agent infrastructure.

  How to use it: Upgrade to v0.26.7; run an admission, recovery, or cleanup flow and check that the verdict behavior is unchanged.

  Tip: In recovery tests, verify that both the duration and verdict remain the same as in the previous version.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Managed Runtime no longer displays internal instructions

  The `record_artifact` directive has been removed from Managed Runtime worker results. The result returned to the user no longer carries this internal recording instruction unnecessarily.

  Usage details

  Who can use it: Qwen Code users who use Managed Runtime.

  How to use it: Run a Managed Runtime task and use the result text directly. If artifact recording is needed, follow the product’s actual artifact flow separately.

  Tip: Check the Managed Runtime recording list instead of assuming an artifact was recorded based on the result text.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Managed turn submission and migration preparation involve less lock contention

  Admission and migration preparation for Managed Agent turn submission have been decoupled from the tenant placement lock where safe. This reduces wait times without changing result behavior.

  Usage details

  Who can use it: Deployments that use Managed Agent turn and migration operations.

  How to use it: Upgrade to v0.26.7 and run the turn submission and migration preparation flows in the same session; monitor operation wait times.

  Tip: Measure error rates and wait times together during migration; don’t treat a shorter queue alone as a sign of success.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code

  Monitor and workflow notifications are attributed to the correct turn

  Monitor and workflow completion notifications are now attributed to the turn that recorded them. This lets the active turn’s results reach the next model request sooner.

  Usage details

  Who can use it: Sessions that use Qwen Code monitor and workflow.

  How to use it: Run a monitor or workflow; observe the model request and result generated in the turn that follows the one that recorded the completion notification.

  Tip: When running many background tasks at once, check the session logs to see which turn each notification came from.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Session commands respond with fewer lockups

  Lock usage has been reduced for common Session commands such as changing the model, approving, context compression, canceling tasks, and renaming. This aims to make the same session respond faster.

  Usage details

  Who can use it: Teams using Qwen Code Managed Agent Session.

  How to use it: Upgrade to v0.26.7 and change the model, approve, start context compression, cancel a task, or rename it in Session.

  Tip: In busy sessions, record the time between command submission and the UI response to measure user-perceived latency.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Resident Runtime Session acquisition and release are faster

  Managed Agent resident Runtime Session claim and release flows no longer wait for the tenant placement lock. Session placement and release wait times have been reduced while preserving successful behavior.

  Usage details

  Who can use it: Managed Agent deployments that use Resident Runtime Session.

  How to use it: Update Managed Agent to v0.26.7 and run the resident Runtime Session flow with the same concurrency.

  Tip: When measuring performance gains, record not only the first response but also the total claim-to-release time.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Workspace operations use the tenant placement lock less often

  Normal workspace execution, member removal, and file rewind operations no longer acquire the tenant-wide placement lock. This reduces unnecessary blocking across workspaces.

  Usage details

  Who can use it: Teams running multiple Managed Agent workspaces within a tenant.

  How to use it: Upgrade Qwen Code Managed Agent to v0.26.7 and use concurrent workspace execution, member removal, or file rewind in the normal workflow.

  Tip: For busy tenants, compare wait times before and after the upgrade using the same workload.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 04 Nov
  CapabilityQwen Code Managed Agent

  Local Workspace read, upload, and status checks are faster

  Local Workspace file read, upload, and status checks reuse CLI capability probes while preserving the same local identity. This reduces the cost of repeated local checks.

  Usage details

  Who can use it: Managed Agent setups using Workspace with a local CLI identity.

  How to use it: Upgrade to v0.26.7; run consecutive file read, upload, and status checks with the same CLI identity.

  Tip: Remember that the probe cache must be refreshed when the identity or capabilities change; don’t compare performance using stale results.

  ✓ Verified by official sourceSource: Qwen Code v0.26.7 sürüm notları
• 03 Nov
  Plugin / IntegrationAgentScope

  AgentScope Java supports the JEV structured decision model

  JEV integration has been added to AgentScope Java through `agentscope-extensions-jev`. Developers can use `JevClient` and review components to add structured decisions to tool selection, model routing, execution guardrails, evidence processing, and trajectory evaluation workflows.

  Usage details

  Who can use it: AgentScope developers using Java 17+; JEV API access and separate JEV credentials are required.

  How to use it: Add the `io.agentscope:agentscope-extensions-jev` dependency to the project; set `TYPESAFE_API_KEY` and run the first decision through `JevClient` and `SystemOneRequest`.

  Tip: Don’t interpret JEV probability as universal truth; calibrate thresholds using your own business data and route uncertain results to human review.

  ✓ Verified by official sourceModels: jev-latestSource: Teaching Agents to Decide: From JEV Integration to the Full Harness Decision Flow
• 03 Nov
  Plugin / IntegrationAlibaba Cloud CMS MCP

  Official remote CMS MCP connection for Cloud Monitor released

  Alibaba Cloud CMS MCP exposes Cloud Monitor operations as remote MCP tools; alert, metric, trace, event, entity, Prometheus, and Grafana surfaces can be discovered in the default `core` ToolSet. The same operation contract as the CLI is preserved, without requiring `aliyun cms2` to be installed on the client machine.

  Usage details

  Who can use it: Alibaba Cloud CMS users; teams using Cursor, Claude Code, Lingma, or their own MCP client. China and international endpoints are available; write operations require appropriate RAM permissions.

  How to use it: Add the region-appropriate Streamable HTTP `/mcp` endpoint to your MCP client with `?toolsets=core`; use `core,notification-channel` if needed. Authenticate with OAuth or an existing Alibaba Cloud session, then run a read-only workspace or alert query first.

  Tip: Don’t put an AccessKey in mcp.json; connect with a read-only RAM role first, then enable the required write ToolSets.

  ✓ Verified by official sourceSource: Friday, 21:47, Checkout Is Slow Again: Twenty Minutes of Troubleshooting with CMS MCP

New in this edition · 57

• 04 Nov
  AvailabilityDeepSeek Harness

  DeepSeek Harness v0.2.2-rc.1 released

  The first release candidate in the DeepSeek Harness 0.2.2 series has been released. It covers user and developer changes since v0.2.1-rc.5, including plugins, mobile use, automation, SDK, and desktop workflows.

  Usage details

  Who can use it: Available in the Web interface and desktop app as part of Public preview; desktop distribution is available for macOS and Windows.

  How to use it: To update from source, pull the git repository and run the npm/pnpm installation, or launch it with the npx @deepseek-ai/dsh web command in the README; desktop users can use the installer on the official download page.

  Tip: Because the release candidate may include breaking changes, test it in a separate profile and workspace before using it in production workflows.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notlarıSource: DeepSeek Harness ürün sayfasıSource: DeepSeek Harness masaüstü indirme sayfası
• 04 Nov
  RetiredDeepSeek Harness

  Minimum supported Node.js version raised to 24.9

  The minimum supported Node.js version for DeepSeek Harness has been raised to 24.9; update Node.js before upgrading if you’re using an older runtime.

  Usage details

  Who can use it: DeepSeek Harness installations run from source code or via CLI; Node.js 24.9 and later.

  How to use it: Check the version with node --version. If it’s earlier than 24.9, upgrade Node.js, then reinstall DeepSeek Harness dependencies.

  Tip: Before switching to the candidate version, also verify that Node.js 24.9 is available in CI and remote SSH environments.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  PTC execution architecture and SSH protocol updated

  Native and PTC tool execution have been split into Agent-scoped plugins; PTC is enabled through a configuration overlay. The PTC Node communication channel has moved to birpc, the SSH protocol has been upgraded from 2 to 3, and run_code returns ABORTED when canceled by the caller.

  Usage details

  Who can use it: Developers who distribute their own bootstrap or SSH helper, and SDK/Session consumers using PTC.

  How to use it: Update the self-hosted bootstrap and SSH helper components together, enable PTC through a configuration overlay, and add ABORTED to the SDK/Session cancellation class.

  Tip: Upgrade the protocol on the client and remote helper in the same deployment; do not mix an old helper with a new client.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Session API provides snapshot and continuous event subscriptions

  Session API provides separate interfaces for reading snapshots and continuously subscribing to session event logs, execution status, and event projection data.

  Usage details

  Who can use it: SDK and plugin developers using DeepSeek Harness Session API.

  How to use it: Call the snapshot read endpoint in Session API; then subscribe to the event log and execution status, and associate incoming events with the Session ID.

  Tip: Add reconnection and resume-from-last-event logic to continuous subscriptions to prevent event loss during brief network outages.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  TypeScript and Python SDKs expanded with Web capabilities

  The SDKs support file delivery, questions and callbacks, scheduling, skills, and context capabilities to better match Web. Plan review, goal persistence, and action approvals have also been added; plugin changes take effect when the SDK restarts.

  Usage details

  Who can use it: Developers using the DeepSeek Harness TypeScript and Python SDKs.

  How to use it: Install the latest SDK, enable file delivery or question callbacks, and use the plan/action approval flow; restart the SDK process after changing a plugin.

  Tip: To compare Web and SDK behavior, run the same session plan in both clients and test callback and approval records.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Web Fetch no longer rejects proxy and internal network addresses

  Web Fetch can also read non-public addresses and allows proxy or internal network use. Since restricting network access is left to the deployment, security policies must be configured separately.

  Usage details

  Who can use it: Self-hosted deployments or deployments behind a proxy that use Web Fetch; internal network addresses may be accessible.

  How to use it: Configure the deployment’s network egress policy before using Web Fetch, then test an allowed internal network or proxy address.

  Tip: Don’t leave internal network access enabled by default; implement allowlist, DNS rebinding, and SSRF checks.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Tools can be enabled based on step conditions

  Tools can now be enabled conditionally at the step level; model notification, PTC context, and execution permission follow the same decision. Plugin developers need to migrate the tools.resolve, tools.executable, and prompt-merging integrations.

  Usage details

  Who can use it: DeepSeek Harness plugins and Agent tools; a new tool resolution contract for developers.

  How to use it: In the plugin, tie tool resolution to the step condition, update the tools.resolve and tools.executable integrations, and test model notification and execution permission together.

  Tip: For conditional tools, restrict the tool schema shown to the model using the same condition as execution.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Behavior for continuing after skipped questions and partial skill discovery fixed

  It is now clear that when a user skips a question, the model should continue without repeating that question, and that skipping does not count as permission. If a skill root cannot be read, skills in healthy roots continue to be used, and incomplete discovery is reported.

  Usage details

  Who can use it: DeepSeek Harness question flows and profiles using multiple skill roots.

  How to use it: Choose skip on a question card and observe the flow continue; temporarily make one skill root inaccessible to verify that the remaining skills are listed.

  Tip: Treat a skipped question as missing a decision rather than as permission, and use a separate approval policy for sensitive actions.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Caching and background notifications improved in Auto review

  Auto review makes better use of cached historical context between consecutive tool calls. When a background job completes, small outputs are included directly in the notification, and duplicate notifications for previously read results are reduced.

  Usage details

  Who can use it: DeepSeek Harness Auto review and background tasks.

  How to use it: Run consecutive tool calls with Auto review; when a task completes, open the small output in the notification and check that previously read results are not repeated.

  Tip: For repeated tool calls, preserve the same context and avoid starting unnecessary new sessions to increase cache gains.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  SettingDeepSeek Harness

  Auto permission and approval policy is preserved on shutdown

  When the app closes, the Auto session's permission and approval policy no longer accidentally switches to Full access; the saved policy is preserved.

  Usage details

  Who can use it: Auto sessions and saved permission/approval policies.

  How to use it: Set the permission policy in an Auto session, close and reopen the app, then verify the previous policy value in the session settings.

  Tip: Check that the permission policy has the expected value before restarting automated tasks.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  AgentDeepSeek Harness

  Delegation tools persist across profile switches, and goals can be resumed from the Web

  A bug that caused delegation tools to disappear after switching profiles and returning was fixed, and delegation now respects the current Agent tool restrictions. Blocked goals can be resumed from the Web target bar; if the turn limit has been reached, increase it first.

  Usage details

  Who can use it: DeepSeek Harness Agent profiles, delegation tools, and the Web target bar.

  How to use it: Switch profiles and return, then try the delegation tool again; resume the blocked goal from the Web target bar or increase the exhausted turn limit first.

  Tip: Give subagents only the tools required for the goal, and set a realistic turn limit from the outset.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Experimental Developer tools Statistics plugin adds a metrics view

  The experimental Developer tools — Statistics plugin shows context, cache, model, and tool times by turn and step.

  Usage details

  Who can use it: Experimental developer tool enabled from the plugin manager.

  How to use it: Enable the Developer tools — Statistics plugin in the plugin manager and open the turn or step metrics for the session.

  Tip: When diagnosing latency, compare model, cache, and tool times separately instead of looking at the total time.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  File paths and Markdown links now open more accurately

  Local tools expand a leading ~ or ~/ to the Host home directory; relative file links in chat use the session working directory at the time of the click. A bare ~ is now plain text in Markdown, and Windows backslashes and punctuation in bare URLs are also handled correctly.

  Usage details

  Who can use it: DeepSeek Harness local file tools, Markdown links, and Windows paths.

  How to use it: Click a file link while in a worktree; try a path starting with ~/ and a Markdown link with Windows backslashes.

  Tip: Because the session working directory can change, document file links with explicit absolute paths when sharing them long-term.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  Plugin / IntegrationDeepSeek Harness

  Fixed plugin ESM resolution and restricted sandbox launch issues

  Plugin dependencies in configuration projects are now resolved using the project’s ESM rules. An issue that prevented restricted sandbox commands from launching when DSH ran with administrator privileges or in Windows Sandbox has been fixed; file write and delete restrictions remain in place.

  Usage details

  Who can use it: Configuration project plugins and Windows Sandbox or administrator-privileged installations.

  How to use it: Launch the plugin configuration project using its own ESM rules; in Windows Sandbox, run a restricted command and separately check the file write and delete policies.

  Tip: Running with administrator privileges does not remove sandbox security boundaries; still test write and delete permissions separately.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  Plugin / IntegrationDeepSeek Harness

  Plugin manager adds rich details and third-party update support

  The plugin manager can display rich-text details, and the page layout and navigation have been improved. Third-party plugin packages can be updated; if a hot update fails, the previous version is restored when possible.

  Usage details

  Who can use it: DeepSeek Harness plugin manager; third-party plugins and packages that support hot updates.

  How to use it: Open the plugin manager, view a plugin’s details, and run the update action. If the hot update fails, check the restoration result.

  Tip: Export your profile or save the release notes before updating third-party plugins.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Copy history as Markdown and preserve replies in collapsed steps

  The full chat history can now be copied as Markdown, with the original log path included when available. When work steps are collapsed, you can choose to preserve all assistant text or show only the final reply.

  Usage details

  Who can use it: DeepSeek Harness session view and work-step view.

  How to use it: Copy the full history as Markdown from the session menu; when collapsing work steps, choose a view that preserves all replies or shows only the final answer.

  Tip: When sharing a bug report, use the Markdown copy that includes the original log path; choose the final-answer view if you only need to read the result.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  Plugin / IntegrationDeepSeek Harness

  SSH fallback and prepare approval added for Git plugin installation

  If HTTPS access to a self-hosted Git plugin fails but SSH access is available, installation switches to SSH. For plugins that include a prepare script, script permissions and retry options are shown.

  Usage details

  Who can use it: DeepSeek Harness plugin manager and self-hosted Git plugins.

  How to use it: Install the Git plugin from the plugin manager; if HTTPS fails, have your SSH key ready, then grant permission at the prepare script prompt and choose to retry.

  Tip: Test HTTPS and SSH access to self-hosted repositories in advance; grant permission to prepare scripts only when you trust their source.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  SettingDeepSeek Harness

  Global and project AGENTS.md budgets are now separate

  Global and project AGENTS.md instructions now use separate maxBytes budgets; project instructions no longer displace global guidance. The combined message total from both chains can now reach up to twice the previous configured limit.

  Usage details

  Who can use it: DeepSeek Harness workspaces that use AGENTS.md.

  How to use it: Keep global and project AGENTS.md files in their respective directories; review the existing maxBytes configuration and monitor the combined context from both instruction chains.

  Tip: Keep important rules concise and prioritized; even with separate budgets, unnecessarily long instructions increase context size.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Images upload before sending and the default limit is now 50 MiB

  An image uploads as soon as it is added to the input field and can be reused when sending. The default single-image size limit has increased to 50 MiB.

  Usage details

  Who can use it: DeepSeek Harness sessions with image support; default 50 MiB limit per image.

  How to use it: Drag or add an image to the input field, wait for the upload to complete, then send the message.

  Tip: For large images, waiting for the preview to be ready before sending reduces the risk of uploading again.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  AgentDeepSeek Harness

  Experimental Auto review support added to headless mode

  Experimental Auto review is now available in headless runs.

  Usage details

  Who can use it: DeepSeek Harness headless runs and the experimental Auto review feature.

  How to use it: Enable Auto review in the headless profile and run a tool task to verify the review result in the logs.

  Tip: Try headless automatic reviews with read-only tools first; add production write permissions later.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Web Search and Fetch tools are unavailable without credentials

  The model’s Web search tools are hidden if the required credentials aren’t available; the new Web API-based Fetch tools are also made available based on credential readiness.

  Usage details

  Who can use it: DeepSeek Harness profiles that use Web Search and Web Fetch and have credentials configured.

  How to use it: Enter the required API credentials in Settings → Web Search and verify the connection; then check the tool list in a new session.

  Tip: Check the settings validation first to distinguish missing credentials from a lack of tool support in the model.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Switching to a smaller-context model and prefix caching are safer

  Before switching to a model with a smaller context window, history is compressed using the old model; if compression fails, the switch is halted and the previous request header is preserved. If the system prompt has not changed, new requests reuse the existing context prefix cache.

  Usage details

  Who can use it: DeepSeek Harness sessions that use model switching and session context.

  How to use it: Select a model with a smaller context window from the model selector and wait for compression to finish; allow the cache to be reused for consecutive requests without changing the system prompt.

  Tip: In long sessions, switch models at a point that leaves time for compression rather than immediately before a critical step.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  AgentDeepSeek Harness

  User steering can immediately stop waiting for background output

  When a user instruction arrives while waiting for background task output, the wait ends immediately and returns the current output; the background task continues running.

  Usage details

  Who can use it: DeepSeek Harness background tasks and user steering.

  How to use it: Send an instruction while a background task is running; receive the current output and check that the background task continues without stopping.

  Tip: If you don’t want to stop a long-running task, send the new request as steering rather than using the cancel action.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Diagnostic report added for desktop startup failures

  For desktop startup failures, a diagnostic report can be obtained that includes a limited history of recent plugin operations.

  Usage details

  Who can use it: DeepSeek Harness desktop app; users experiencing startup errors.

  How to use it: When the desktop app fails to start, open the diagnostic report option and attach the plugin operation history from the report to your support request or bug report.

  Tip: Before sharing the report, check that it does not contain an API key or sensitive file paths.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Messages and drafts are now finalized only after Host acceptance

  A message isn’t displayed and the draft isn’t cleared until it’s accepted by Host. The input field is read-only during submission; if an error occurs, the draft, quote, and attachments are preserved.

  Usage details

  Who can use it: DeepSeek Harness Web and desktop session input field.

  How to use it: Send a message and wait for Host acceptance. If submission fails, retry with the draft, quote, and attachments in the input field.

  Tip: When working with large attachments, don’t close the window without checking that the content was preserved after a submission error.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  SettingDeepSeek Harness

  Improved model picker and provider-specific model ordering

  Configured models are selected by default when the model picker opens; adding and removing candidates, and selecting all visible candidates, are supported. Models from the DeepSeek and pi-ai providers can be reordered by dragging or using the arrow keys.

  Usage details

  Who can use it: DeepSeek Harness model picker; DeepSeek and pi-ai provider configurations.

  How to use it: In Settings → Models, select or remove candidates; drag a model in the DeepSeek or pi-ai provider list, or use the arrow keys, then save.

  Tip: Move your most-used model to the top of the list to reduce the chance of selecting the wrong provider and avoid unnecessary clicks.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Node PTC adds VM Script mode and optional state persistence

  Node PTC now includes VM Script mode, which supports top-level await, dynamic imports, and expression results. State persistence across calls is optional; by default, execution resets on every call.

  Usage details

  Who can use it: DeepSeek Harness plugins and SDK flows that use Node PTC.

  How to use it: Enable VM Script mode in the Node PTC configuration; also enable state persistence if you need to share variables and imports across calls.

  Tip: Enable state persistence only in trusted, short-lived sessions; the default behavior of independent calls is more isolated.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  Plugin / IntegrationDeepSeek Harness

  Custom plugin market migrated to an index-based flow

  Custom plugin market has been renamed Custom plugin index; the discovery page reads user-added index files and excludes custom/private releases. Official bundle lists are exported from the dsh-official-bundles package.

  Usage details

  Who can use it: DeepSeek Harness plugin discovery and custom plugin indexes; integration developers.

  How to use it: Use Custom plugin index in the plugin discovery settings, add the user index file, and check that custom releases aren’t listed; for custom app-boot integrations, switch to the dsh-official-bundles source.

  Tip: List only published, trusted releases in custom indexes; don’t include packages tagged private in the distribution chain.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Custom model capacity and step model routing have changed

  For custom models with unspecified capacity, the default context is set to 1M and the maximum output to 128K. Agent now fixes the step model before preparing the prompt and before the pre-step; the same model is retained for retries, and plugins must use agent.model to select a model.

  Usage details

  Who can use it: Custom OpenAI-compatible endpoints and Agent plugins.

  How to use it: Explicitly configure context and output capacity for a custom model with lower capacity; use agent.model to select a model from a plugin, and don’t switch routes with agent/request.

  Tip: If your endpoint doesn’t support 1M/128K, explicitly set a lower capacity instead of relying on delayed compaction.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Upgraded to pi-ai 1.3.0; malformed tool parameters are rejected early

  The model catalog, endpoints, and reasoning replay rules have been updated with pi-ai 1.3.0. If completed tool parameters are malformed, the response is terminated before any tool runs.

  Usage details

  Who can use it: DeepSeek Harness plugins and providers that use pi-ai routes.

  How to use it: Update the profile that uses pi-ai and verify that tool schemas produce valid JSON parameters; reload the model catalog after the update.

  Tip: Validate tool parameters against the schema before production so early terminations don’t disrupt the user flow.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  RetiredDeepSeek Harness

  Ralph loop and Session search removed from the experimental list

  Ralph loop and Session search are no longer shown in the experimental plugins list. This indicates that the related experimental list and discovery path have been removed.

  Usage details

  Who can use it: DeepSeek Harness experimental plugin list and profile discovery.

  How to use it: Open the experimental list in the plugin manager and check that Ralph loop and Session search are no longer listed; update profiles that use them according to the migration notes.

  Tip: Export profiles that depend on removed experimental items and identify an alternative workflow before upgrading.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  sdk-minimal keeps recovery records for short sessions

  sdk-minimal writes a session recovery record before a model request or tool run. If the record cannot be written, the operation does not proceed; this makes it safe to resume interrupted short sessions.

  Usage details

  Who can use it: DeepSeek Harness sdk-minimal configuration and sessions created with the SDK.

  How to use it: Start a session with sdk-minimal, configure storage where the recovery record can be written, then run a model or tool call.

  Tip: Store recovery records in persistent, backed-up storage rather than a temporary directory.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Long Search API summaries are prioritized for the model

  Instead of the short excerpts on cards, the Search function prioritizes long Search API summaries of up to 800 characters per source by default for the model.

  Usage details

  Who can use it: DeepSeek Harness searches using Search API; the default source summary limit is 800 characters.

  How to use it: Enable Web search and run a search task; inspect the long source summaries in the model context separately from the short search card shown to users.

  Tip: For research involving multiple sources, preserve the 800-character summaries and use only verified claims in the final answer.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  APIDeepSeek Harness

  Session Projection and storage compatibility changes

  The Session Projection registration API is now asynchronous, and plugins must retain a handle at startup. JSON and SQLite storage can read older versions explicitly declared as compatible and write in the current format.

  Usage details

  Who can use it: Session Projection plugins and self-hosted installations using JSON/SQLite.

  How to use it: Wait for the asynchronous Projection registration result when initializing the plugin and store the handle; explicitly specify a compatible older version for storage.

  Tip: Back up the database before upgrading storage, and enable only documented compatibility versions.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Add chat text to the input field as a quote

  Text from chat can now be transferred to the input field as a quote. After submission, the quote is displayed as a brief preview.

  Usage details

  Who can use it: Chat input field in the DeepSeek Harness web and desktop interfaces.

  How to use it: Select the relevant text in a chat message and transfer it to the input field as a quote; check the compact preview before sending.

  Tip: In code reviews, quote only the necessary lines to reduce context and token usage.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  Plugin / IntegrationDeepSeek Harness

  The subagent_oar plugin supports local Codex and Claude CLI tasks

  The experimental subagent_oar tool plugin adds the ability to delegate independent tasks to local Codex or Claude CLI processes.

  Usage details

  Who can use it: Available as an experimental plugin in DeepSeek Harness profiles; on machines with local Codex or Claude CLI installed.

  How to use it: Enable the experimental subagent_oar plugin in the plugin manager, configure local CLI credentials, and delegate the independent task to this tool.

  Tip: Keep workspace permissions as narrow as possible before granting local CLIs access to files and commands.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Delivery cards now include file references from the entire turn

  File links in delivery cards and closing responses now include deliveries and writes made after the response. It’s clarified that Agent must provide the file link before waiting for a response from the user.

  Usage details

  Who can use it: DeepSeek Harness file deliveries and closing responses.

  How to use it: Create or write a file during a turn, then open the delivery card and the file link in the closing response to check the final file.

  Tip: In workflows that wait for user approval, provide the file link first; don’t leave it until the next turn.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Trajectory preserves actual output from failed tools and legacy code highlighting

  The actual text output from failed tools is preserved in the Trajectory list, details, and copied content. TypeScript and Python syntax highlighting is restored for affected legacy run_code records.

  Usage details

  Who can use it: DeepSeek Harness Trajectory view and historical run_code records.

  How to use it: After a failed tool call, open the Trajectory details and copy output; select the language used in a legacy run_code record and check the highlighting.

  Tip: When debugging, rely on the tool’s actual return text, not a generated error summary.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Trajectory adds a raw JSON view for live tool results

  The Trajectory results page can show the raw JSON of the live tool response above the processed view.

  Usage details

  Who can use it: DeepSeek Harness Trajectory results page.

  How to use it: After a tool call completes, open the Trajectory results page and expand the raw JSON view above the processed result.

  Tip: When diagnosing an integration error, compare raw JSON field names instead of processed text.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  SettingDeepSeek Harness

  WebSocket compression is now enabled by default for remote connections

  WebSocket compression is enabled by default to reduce the transfer of large histories over remote connections. ETag validation for unchanged web static assets also prevents them from being downloaded again with a 304 response.

  Usage details

  Who can use it: Remote DeepSeek Harness connections; compression can be disabled with the websocketCompression setting.

  How to use it: Start the remote connection as usual; to disable compression, set websocketCompression to false in the configuration.

  Tip: Keep default compression enabled on slow networks; on CPU-constrained local networks, consider disabling it only after measuring.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  AvailabilityDeepSeek Harness

  Desktop update flow improved on Windows and macOS

  Windows installer progress and window interactions have been fixed; the app publisher and names in the uninstall list are now displayed correctly. macOS updates now handle exiting full screen, workspace recovery after a failed installation, and in-window notifications that match the app theme more effectively.

  Usage details

  Who can use it: DeepSeek Harness desktop app; Windows and macOS.

  How to use it: Use the in-window flow when an update notification appears in the desktop app; on macOS, exit full screen before confirming the update and use the retry option if it fails.

  Tip: Save the status of open tasks before a desktop update, and make sure the update window is visible in full-screen apps.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notlarıSource: DeepSeek Harness masaüstü indirme sayfası
• 04 Nov
  AvailabilityDeepSeek Harness

  Six new interface languages added

  Text and localization configurations have been added for Japanese, Spanish, French, German, Russian, and Brazilian Portuguese.

  Usage details

  Who can use it: DeepSeek Harness interface; Web and desktop clients with the relevant localization configuration.

  How to use it: Choose one of the supported languages from the language selector in Settings, then reopen the interface to apply the localization.

  Tip: For team use, agree on a shared language to avoid confusion over screenshots and menu names.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Workspaces and sessions with the same name are easier to distinguish in the sidebar

  When selecting workspaces with the same name, the shortest distinguishing path suffix is shown. The working directory is now visible on session cards, and sidebar search matches session ID fragments regardless of case.

  Usage details

  Who can use it: DeepSeek Harness workspace selector, sidebar, and session search.

  How to use it: Open workspaces with the same name in the selector; hover over a session in the sidebar or type part of a session ID in the search box.

  Tip: Adding the last few characters of a session ID to a team note makes it easier to find sessions with similar names.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Connection downtime and Trajectory message order fixed

  The chat timer pauses when the connection is lost or being re-established, then resumes upon reconnection. After a stop, subsequent routing and follow-up messages are preserved in Trajectory in log order, before the assistant output.

  Usage details

  Who can use it: DeepSeek Harness web and desktop sessions; Trajectory history.

  How to use it: Disconnect and reconnect an active session, then check the timer; review follow-up messages after stopping in Trajectory.

  Tip: For long tasks with network interruptions, use the interface timer—not wall-clock time—to measure actual working time.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Pan large images and deliver original files by dragging

  Images that exceed the viewport can now be panned by dragging. Original files can be delivered using in-app or desktop drag and drop.

  Usage details

  Who can use it: DeepSeek Harness Web and desktop image previews.

  How to use it: Drag an image that exceeds the viewport in the preview; drag and drop the original file into the app or onto the desktop.

  Tip: For analysis, deliver the original file by drag and drop instead of using a reduced preview.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Experimental personalized plugin recommendations added

  The experimental personalized recommendations plugin can show cards under relevant responses for enabling an official plugin or creating a personalized page.

  Usage details

  Who can use it: DeepSeek Harness users who enable the experimental Personalized recommendations plugin.

  How to use it: Enable the Personalized recommendations plugin in the plugin manager; use the enable or page-creation cards under relevant responses.

  Tip: Review which plugins are trusted before enabling recommendation cards as a team standard.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  File search and Word delivery work more efficiently

  File searches with a fixed directory prefix are faster because irrelevant directories are skipped. Routine Word edits are delivered after content and formatting are saved and verified, without unnecessary ZIP, XML, or metadata inspection.

  Usage details

  Who can use it: DeepSeek Harness file tools and Word tasks.

  How to use it: Start a file search with a specific directory prefix; after editing a Word document, check the save and formatting verification results, then use delivery.

  Tip: In large repositories, limit searches to the working directory or a fixed prefix whenever possible.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  File tab position and dark theme settings navigation fixed

  The scroll position is preserved in sidebar file tabs after selecting, opening, or closing a tab. The hover background for settings navigation is now visible in dark theme.

  Usage details

  Who can use it: DeepSeek Harness sidebar file tabs and dark theme settings.

  How to use it: Open multiple file tabs in the sidebar and close one; then hover over Settings navigation in dark theme.

  Tip: When reviewing multiple files, use preserved scroll position while closing tabs to avoid searching the file list again.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Input and session statistics expanded

  Truncated statistics in the input field can now be viewed in full by hovering over them. Session statistics also show the number of turns and steps.

  Usage details

  Who can use it: DeepSeek Harness input field and session statistics panels.

  How to use it: Hover over a truncated statistic; open session statistics to review the number of turns and steps.

  Tip: When assessing token usage or tool duration in long tasks, record the number of turns and steps at the same time.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Native macOS folder picker available on first launch

  Fixed an issue that caused the native macOS folder picker to appear inactive on first launch and reject folder clicks.

  Usage details

  Who can use it: DeepSeek Harness macOS desktop app.

  How to use it: When adding a workspace, open the folder picker, click a folder, and confirm that the selection is accepted.

  Tip: If the folder does not appear on first launch, verify that the latest candidate build is installed instead of restarting the app.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Regular conversation branches are no longer counted as subagents

  Regular conversation branches run independently and are no longer mistakenly classified as subagents; the creator relationship is preserved in actual subagent sessions.

  Usage details

  Who can use it: DeepSeek Harness session branches and actual subagent sessions.

  How to use it: Create a regular branch from a session and observe it running independently; when you create a subagent, check that the parent session relationship is preserved.

  Tip: Use regular branches for independent experiments, and subagents for delegated tasks and linked work.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Right sidebar tabs can be closed with the middle mouse button

  Right sidebar tabs can be closed with the middle mouse button; closing a background tab does not change the current selection, and the initial tab is preserved.

  Usage details

  Who can use it: The right sidebar in the DeepSeek Harness desktop and Web interfaces.

  How to use it: Middle-click the tab you want to close in the right sidebar; the initial tab cannot be closed.

  Tip: When many files or results tabs are open, use the middle mouse button to quickly close only the unnecessary tabs.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Completed automation tasks can be cleared with a single confirmation

  The automation tasks page now includes an option to clear all inactive tasks with a single confirmation while preserving running tasks.

  Usage details

  Who can use it: DeepSeek Harness automation tasks page.

  How to use it: Open the automation tasks page, select the action to clear completed or failed tasks, and confirm; ongoing tasks are preserved.

  Tip: Export the output of failed tasks before clearing; the action removes only inactive tasks.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Interactions in question and approval cards clarified

  For multiple-choice questions, all recommended options are selected by default; long question and option descriptions can be expanded and collapsed inline. Approval messages clearly indicate whether permission is actually being requested from the user.

  Usage details

  Who can use it: DeepSeek Harness question cards, tasks with options, and approval flows.

  How to use it: Open a multiple-choice card and check that recommendations are selected; expand a long description and read the permission wording in the approval message before proceeding.

  Tip: Review automatically selected options individually before submitting; a recommended selection does not imply permission.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Overwritten local Markdown images now display the latest version

  After a local Markdown image is overwritten, new messages and newly opened previews no longer use the stale image from the cache.

  Usage details

  Who can use it: DeepSeek Harness local Markdown image previews.

  How to use it: Replace the local Markdown image, then open a new message or preview; verify that the latest file is displayed.

  Tip: In image-generation workflows, open a new preview after each change to avoid accidentally using a stale cached image.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  CapabilityDeepSeek Harness

  Web Fetch can read pages with mixed UTF-8 declarations

  Web Fetch can now read valid pages whose encoding declaration contains multiple labels alongside UTF-8. This eliminates unnecessary read errors on pages with multiple encoding declarations.

  Usage details

  Who can use it: Web Fetch tool and web pages serving valid UTF-8 content.

  How to use it: Enable Web Fetch and request the URL of a page with mixed encoding declarations; check that the returned text is not corrupted.

  Tip: If the encoding issue persists, also verify the page’s raw response headers and content encoding.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları
• 04 Nov
  FeatureDeepSeek Harness

  Answered question groups remain in chat bubbles

  Answered question groups are displayed as message bubbles. Responses to Native and PTC calls remain visible even when work steps are collapsed.

  Usage details

  Who can use it: DeepSeek Harness sessions and question-and-answer tool calls.

  How to use it: Answer when a tool call asks a question, then collapse the work steps and check that the answered question group remains in a chat bubble.

  Tip: Keep the session record instead of deleting question bubbles to preserve decision history in multi-step tasks.

  ✓ Verified by official sourceSource: DeepSeek Harness v0.2.2-rc.1 resmî release notları