每一个发布版本

内容直接来自 GitHub 上已发布的版本。新版本会自动出现在这里，无需改动本站。

更新日志取自各版本的 GitHub 发布说明，因此以英文呈现。

Added

Connect Grok Build and Kimi Code accounts through their official sign-in flows, with International and Mainland China regions for Kimi.
Connect GLM (Z.AI and Zhipu AI), Kimi For Coding, MiniMax international and mainland China, GitHub Copilot, and GitLab Duo coding plans from Accounts through the installed OpenCode CLI.
Provider tools shows installed and latest versions for Codex, Claude Code, Cursor, OpenCode, Grok, and Kimi, with Update all and an optional automatic update check every six hours.
The desktop's remote bridge streams chat, tool activity, diffs, git progress, and terminals to the DJL iPhone app with lower latency and reports when the Mac is offline.

Fixed

The desktop window no longer opens blank when a remote font stylesheet stalls. Fonts are bundled, and the window appears once the app has rendered.

Changed

DJL now uses your separately installed OpenCode CLI for login, model discovery, and chat instead of a bundled copy. Existing conversations migrate on first resume, and copying older DJL-only logins to OpenCode is an explicit action in Accounts.

在 GitHub 上查看

Added

Added dedicated China download choices for the matching macOS and Windows installers.
Added anonymous website visit, download, and first-launch counts without collecting user content.

Fixed

Prevented incomplete Hong Kong mirror uploads from advancing the stable desktop update feed.

Changed

Desktop updates now prefer the Hong Kong Alibaba OSS mirror and retry through GitHub for eligible network and download failures.
Stable releases now publish immutable installer files and checksums to the Hong Kong mirror before advancing updater metadata.

在 GitHub 上查看

Added

Windows installers are now Authenticode signed and RFC 3161 timestamped by Anthony Su through Microsoft Artifact Signing.

Changed

GitHub Actions now authenticates to Azure with secretless OIDC and stops the release if the Windows signer or timestamp is invalid.
Download, security, and release guidance now identifies signed Windows installers as the required production policy.

在 GitHub 上查看

Added

Added guided first-run setup and multilingual recommendations for choosing between local and API models.
Added a grounded local-model workflow that connects visible tools, project memory, and preparation state before changes are proposed.

Fixed

Improved tool selection and tool-call recovery across local and hosted models, including DeepSeek text-form tool calls and web research.
Fixed streamed assistant responses so partial text remains complete and correctly ordered while a turn is running.
Hid desktop update and task-panel controls when those actions are unavailable.

Changed

Made switching providers and models clearer in the composer and settings, with stronger guidance for model capabilities and tradeoffs.
Improved local-model runtime lifecycle handling and grounded-work reliability.

在 GitHub 上查看

Added

Added a replayable first-run tutorial for Work, Projects, local AI, and Settings.
Added one-click local AI setup by purpose—General, Documents, Reasoning, or Coding—with hardware-aware recommendations, readiness checks, and automatic fallback to a safer model in the same category.
Added Get Started, Guide, and Changelog pages to the DJL website with release-aware desktop downloads.

Fixed

Fixed local AI hardware detection on Windows and translated Apple Silicon builds so model recommendations use the host CPU, GPU memory, system memory, and disk capacity more reliably.
Fixed local AI setup continuity when a recommended model cannot load, including reuse of the active setup while DJL downloads the fallback model.
Fixed AI detector cache invalidation so updated calibration and model-output settings cannot reuse stale analysis results.
Fixed the desktop updater’s recovery path when version context is missing or a previous update action failed.

Changed

Kept the desktop update control available as a consistent place to check, download, and install new DJL versions.
Expanded AI detector evaluation coverage, metrics, and model-license evidence across English and Chinese benchmark sources.

在 GitHub 上查看

Added

Added streamlined local-model setup with hardware-aware recommendations, measured generation speed, and installed-model management for Ollama and LM Studio.
Added a sidebar update indicator when a new DJL desktop version is ready.
Added verifiable build provenance for desktop installers.

Fixed

Fixed local-model capability routing so chat-only models receive no work tools while capable models continue through the OpenCode tool harness.
Fixed LM Studio context handling to respect each loaded model's effective context window, and stopped overriding Ollama's own context default.
Fixed local runtime discovery, startup, provider labels, and model inventory synchronization.
Fixed duplicate document conversions that could start concurrently while resuming work after a restart.
Fixed desktop release validation to enforce the current 13-asset public release contract in required CI.

Changed

Simplified local-model installation to one recommended action, with additional models and runtimes available under advanced choices.
Updated desktop download and release automation to advertise the current published version and validate native bundles consistently across CI and release builds.

在 GitHub 上查看

First public release of the DJL desktop app.

DJL is a local-first desktop workspace for coding agents. It brings chats, terminals, browser previews, diffs, Git operations, provider sessions, and task handoffs into one focused application.

Highlights

Use your existing Codex, Claude Code, Gemini, OpenCode, Cursor, Grok, Kilo Code, and Pi accounts.
Run parallel tasks across projects and isolated Git worktrees.
Keep conversations, terminals, browser previews, diffs, and agent activity side by side.
Review changes and complete branch, commit, push, and pull-request workflows in place.
Pair the optional DJL iPhone app with the desktop over an end-to-end encrypted remote channel.

在 GitHub 上查看