v2.0.1

MuxLayer 2.0.1 — DeepSeek image understanding

This patch release adds model-specific image support for DeepSeek without changing the behavior of its text-only models.

Highlights / 重点

• New vision model / 新视觉模型 — deepseek-v4-flash-vision-exp is available in the built-in provider catalog and accepts image input through Chat Completions, Responses, and Anthropic-compatible routes.
• Strict model boundary / 严格模型边界 — deepseek-v4-flash and deepseek-v4-pro remain text-only. MuxLayer does not pass images to them and keeps its explicit degradation notice on conversion paths.
• End-to-end vision routing / 完整视觉路由 — Image requests can promote a DeepSeek text model to the vision model, and that promotion is preserved through native Chat, Responses, and Anthropic pass-through.
• Verified upstream behavior / 已验证上游行为 — Live DeepSeek API checks confirmed the vision model accepts images while Flash and Pro reject them.
• Bilingual documentation / 双语文档 — English and Chinese guides and reference tables now describe the exact per-model behavior.

v2.0.0

MuxLayer 2.0.0 — Brand migration with automatic legacy data adoption

MuxLayer is now the primary product identity: the local model control layer for coding agents.

Highlights / 重点

• MuxLayer brand migration / 品牌迁移 — README, documentation, website, app title, GitHub metadata, and release presentation now use MuxLayer.
• Automatic legacy data adoption / 老用户自动接管 — Existing desktop installations keep the legacy updater identity and application data directory so upgrades do not lose databases or tokens.
• Headless data compatibility / 无界面模式兼容 — New CLI installations use ~/.muxlayer; an existing ~/.agentgate database or token is detected and reused automatically.
• Environment aliases / 环境变量别名 — MUXLAYER_DB_PATH and MUXLAYER_TOKEN are preferred; AGENTGATE_DB_PATH and AGENTGATE_TOKEN remain supported.
• Rollback-safe transition / 可回滚迁移 — No old directory is deleted, moved, or overwritten in this release.
• Canonical brand surface / 正式品牌入口 — The MuxLayer logo, guide filenames, comparison URLs, sitemap, and SEO pages are now canonical; old AgentGate guide URLs redirect to their MuxLayer equivalents.
• Official Homebrew cask / 正式 Homebrew 安装入口 — New installations use brew install --cask dengmengmian/tap/muxlayer; the legacy agentgate cask remains upgrade-compatible.
• Native app icon refresh / 原生应用图标更新 — macOS, Windows, Linux, Android, and iOS package assets now use the MuxLayer icon.
• Verified provider catalog / 已验证供应商目录 — Product surfaces consistently describe the built-in catalog as supporting 25+ model providers.

The legacy technical identifiers remain intentionally compatible for this release cycle: agentgate-serve, the agentgate virtual model, the updater bundle ID, and legacy data paths and environment variables. The agentgate Homebrew cask remains available as an upgrade-compatible alias while muxlayer is the primary cask.

MuxLayer 1.6.3 — Outbound proxy & routing that actually matches the request

AgentGate 1.6.3 — Outbound proxy & routing that actually matches the request

What's new

• Outbound HTTP proxy in Settings. Send provider API calls through a local http:// or https:// proxy (Clash / V2Ray). Loopback stays direct. Leave the switch off to keep using HTTP(S)_PROXY from the environment. An empty URL does not apply the UI proxy.
• Routing conditions apply on Chat, Messages, and Gemini. Input size, images, tools, system keywords, and model-name match now evaluate on those routes the same way they already did for Codex Responses. Existing profiles stay as they are; previously ignored conditions now fire.
• Gemini follows the daily budget gate. generateContent is subject to the same notify / block / force-cheapest policy as other routes.
• Less work on the request path. Budget-off skips the lifetime stats scan; daily spend is a cached SUM. Each route parses the request JSON once. SSE splitting and session lookup are cheaper. Overview and footer poll less often.

本次更新

• 设置里可配出站 HTTP 代理。 访问上游供应商的请求可走本机 http:// / https:// 代理（Clash / V2Ray）；回环地址仍直连。关闭开关则继续尊重环境变量 HTTP(S)_PROXY。地址为空时不会走这里的代理。
• Chat / Messages / Gemini 也会看路由条件。 输入长度、是否有图、是否有工具、系统关键词、模型名匹配，与 Codex Responses 同一套规则。已有路由配置不用改，只是以前被忽略的条件现在会生效。
• Gemini 走日预算闸。 generateContent 与其它入口一样受仅提醒 / 拦截 / 强制最便宜策略约束。
• 热路径更轻。 预算关闭时不再扫全量统计；日花费用短缓存 SUM。每条路由的请求 JSON 只解析一次。SSE 拆行与会话查找更便宜。概览和页脚轮询更慢。

v1.6.2

AgentGate 1.6.2 — First-run clarity & failure timeline

What's new

• Honest first-run path. Quick Setup only marks complete when provider, gateway, client apply, and probe all succeed; failure shows one primary CTA (edit key / retry / providers / clients / logs). Success offers copyable launch commands for wired clients only.
• Dashboard no longer pretends every local config is ready. “Ready for first request” requires a usable API key and has_agentgate clients — not merely config_exists on disk.
• Stream failure timeline. Failed request details show start → route → upstream status → last SSE event → failover chain → end reason, built from existing trace_json / sse_events.
• Session UX refresh. Request vs Session switch is primary; session mode uses lighter filters and card list with open-conversation as the main action.
• Less UI noise. Local models open from a dialog; provider refiner tips use plain language in expanded details; MCP “more” menu is no longer clipped by header overflow.

本次更新

• 首装路径更诚实。 快速配置仅在供应商、网关、客户端 apply、连通探针都成功时算完成；失败只给一个主操作。成功页仅对已接入客户端展示可复制启动命令。
• 概览不再误报就绪。 「可以发起第一次请求」要求可用 API Key 且客户端已 has_agentgate，不再把本机有配置文件当成已接入。
• 流式失败时间轴。 失败请求详情展示：开始 → 选路 → 上游状态 → 末次 SSE → 故障转移 → 结束原因（用现有 trace / sse 字段）。
• 会话交互重做。 请求/会话切换置顶；会话模式筛选更轻，卡片列表主按钮打开对话。
• 界面减噪。 本地模型改为弹窗；供应商兼容提示改成人话并收进详情；MCP「更多」菜单不再被页头裁切。

v1.6.1

AgentGate 1.6.1 — Responses pass-through fixes

What's new

• Pass-through no longer drops tools. Codex gpt-5.6+ ships tool definitions inside input.additional_tools; native Responses pass-through forwarded the raw body without hoisting them, so upstreams saw no tools and the model wrote fake tool calls into its answer.
• Pass-through records tokens and cost again. Responses-shaped usage (response.usage with input_tokens / output_tokens) is now parsed alongside the Chat Completions shape.
• Capability gate before pass-through. When the target model or a requested custom tool is outside what the upstream Responses API accepts, the request falls back to protocol conversion instead of failing upstream.
• DeepSeek guidance. The Codex + DeepSeek guides explain, with measured evidence, why AgentGate keeps protocol conversion as the default instead of connecting directly to DeepSeek's Responses API.

本次更新

• 直通不再丢工具。 Codex gpt-5.6+ 把工具定义放在 input.additional_tools 里，而原生 Responses 直通转发 body 原文时没有做提升，上游收不到任何工具，模型只能在正文里编造工具调用。
• 直通恢复 token 与成本统计。 Responses 形态的 usage（response.usage 下的 input_tokens / output_tokens）现已与 Chat Completions 形态一并解析。
• 直通前的能力判定。 目标模型或请求中的自定义工具超出上游 Responses API 的接受范围时，自动回落到协议转换，不再把注定失败的请求发出去。
• DeepSeek 说明。 Codex + DeepSeek 教程补充实测依据，说明为什么默认保持协议转换、而不是直连 DeepSeek 的 Responses API。

v1.6.0

AgentGate 1.6.0 — Session trust, budget gate, local models

What's new

• Session affinity. Multi-turn chats stick to the same upstream provider unless failover or budget override needs another path.
• Daily budget gate. Cap daily spend with notify-only, hard block, or force-cheapest routing for new requests (streams already in flight continue).
• Request Diff + redacted repro export. Compare raw vs converted bodies in the log drawer; copy or download a redacted package for GitHub issues.
• Local models. Scan Ollama / LM Studio (and common local ports) from Providers and add them in one click. Guide: use local models.
• Editor clients. Clearer Gateway connection info and a guide for Cursor / Continue / Cline: editor guide.
• Onboarding & clients. Setup completion requires a real provider check; Clients can show running-process status.

本次更新

• 会话亲和。 多轮对话默认粘在同一上游；失败转移或预算强制最便宜时仍可换路。
• 日预算闸。 支持仅提醒 / 拦截 / 强制最便宜；只限制新请求，不中途掐断已在流式中的响应。
• 请求 Diff + 脱敏复现包。 日志详情对比原始与转换 body，可复制或下载脱敏包用于提 issue。
• 本地模型。 供应商页扫描 Ollama / LM Studio 等本地端点并一键添加。指南：本地模型。
• 编辑器客户端。 网关连接信息更清晰，并提供 Cursor / Continue / Cline 接入指南：编辑器指南。
• 引导与客户端。 完成引导需真实供应商连通；客户端列表可显示进程是否在跑。

v1.5.1

AgentGate 1.5.1 — Kimi K3

What's new

• Kimi K3 models. Built-in catalog adds Platform kimi-k3 and Kimi Code k3 (1M context), plus highspeed coding models. Default model is kimi-k3.
• Correct K3 thinking. K3 uses top-level reasoning_effort: max instead of the K2 thinking parameter.
• Kimi Code keys just work. Paste a sk-kimi-… membership key and AgentGate auto-fills the coding endpoint and model k3.

本次更新

• Kimi K3 模型。 内置目录补齐 Platform kimi-k3、Kimi Code k3（1M 上下文）及高速编码模型；默认模型为 kimi-k3。
• K3 思考参数对齐。 K3 走顶层 reasoning_effort: max，不再误用 K2 的 thinking。
• Kimi Code key 即用。 粘贴 sk-kimi-… 会员 key 时自动填 Code 端点与模型 k3。

v1.5.1

AgentGate 1.5.1 — Kimi K3

What's new

• Kimi K3 models. Built-in catalog adds Platform kimi-k3 and Kimi Code k3 (1M context), plus highspeed coding models. Default model is kimi-k3.
• Correct K3 thinking. K3 uses top-level reasoning_effort: max instead of the K2 thinking parameter.
• Kimi Code keys just work. Paste a sk-kimi-… membership key and AgentGate auto-fills the coding endpoint and model k3.

本次更新

• Kimi K3 模型。 内置目录补齐 Platform kimi-k3、Kimi Code k3（1M 上下文）及高速编码模型；默认模型为 kimi-k3。
• K3 思考参数对齐。 K3 走顶层 reasoning_effort: max，不再误用 K2 的 thinking。
• Kimi Code key 即用。 粘贴 sk-kimi-… 会员 key 时自动填 Code 端点与模型 k3。

v1.5.0

AgentGate 1.5.0 — Keep Awake / 防休眠

What's new

• Keep awake by default on macOS and Windows. AgentGate prevents automatic system sleep for its lifetime while still allowing the display to turn off by default.
• Request-aware control. Optionally keep awake only while AI generation requests are active and during a configurable cooldown (15 minutes by default). Concurrent requests, streaming responses, and disconnects are tracked correctly.
• Fast controls everywhere. Settings show live status and real backend errors; the system tray provides one-click toggles for keep-awake, request-aware control, and keeping the display awake.
• Persistent, immediate state. Configuration survives restarts, mode changes reconcile immediately, and system assertions are released when AgentGate exits.

本次更新

• macOS 和 Windows 默认防休眠。 AgentGate 运行期间阻止系统自动休眠，默认仍允许显示器关闭。
• 请求智能控制。 可选仅在 AI 生成请求进行中及冷却期内保持唤醒，默认冷却 15 分钟；并发请求、流式响应和客户端断开都会正确计数。
• 设置页与托盘快速操作。 设置页显示实时状态和真实系统错误；托盘可一键切换防休眠、请求智能控制和显示器常亮。
• 持久化并立即生效。 配置重启后保留，模式切换立即重建系统申请，AgentGate 退出时可靠释放。

Platform support / 平台支持

• macOS: supported / 支持
• Windows: supported / 支持
• Linux: tray controls are available, but keep-awake is explicitly marked unsupported in this release / 托盘菜单可用，但本版防休眠会明确显示暂不支持

v1.4.13

AgentGate 1.4.13

English section headings are generated automatically. Detailed notes follow the bilingual changelog source.

Improvements / 改进

• Denser provider layout / 更紧凑的供应商布局 —— The Providers page now uses the available desktop width with a responsive two-column grid, flatter information grouping, and a compact bottom-aligned action row. 供应商页改为桌面端双列响应式布局，减少嵌套容器并压缩底部操作区，充分利用横向空间。
• Single first-run provider / 首次初始化仅保留一个供应商 —— New or empty databases now seed only DeepSeek instead of also adding an unused Custom OpenAI Compatible placeholder. Existing provider records are never removed or changed. 新安装或空数据库现在只初始化 DeepSeek，不再额外创建未配置的 Custom OpenAI Compatible；已有供应商记录不会被删除或修改。