Model Adapter & Client Adapter

Gemini thought display mode in Claude:

Changelog

1. 优化 Model Adapter 构架， 也就是说从这个版本开始会增加对模型私域的针对优化能力，以避免切换模型后对上下文的兼容性导致的输入、输出能力不一致问题。现在在同一个session中切换模型也可以获得完全一致的效果（不包括推理能力差距）推荐在智能体应用中只使用主备聚合模式，即一种模型使用多个可个负载KEY提升负载能力的同时对项目思考的智力水平也不会产生大的漂移。

2. 新增 Client Adappter 构架。可以增对 Codex / Claude / Opencode 此类客户端针对性检测和优化，可以面向不同客户端发送不同的思维链摘要区块格式，以达到美观的渲染效果 （前提是extra body 里针对模型开启思考模式）

3. 现在 Deepnow 不仅可以协调9种协议端到端之间相互翻译、透传，还可以针对模型、客户端进行检测和优化，使很多专属 Agent 可以挂载任意模型达到和原生模型一样的显示效果（推理能力除外）

4. Dashbaord 里 "规则化" 配置-> "大模型规则" 新增一个 "思维链构建策略" 可以针对开启思考模式后类 think 文字的显示。各家可能不一样，比如 gemini 是 thought ，deepseek 是 reasoning ，其它的则是 think ，但是 Deepnow 可以把这些不一样统一起来。

5. 增加一个 sse watchdog func

Changelog

1. Model Adapter Architecture Optimization
   Starting from this version, target optimization capabilities for the model's private domain have been added. This prevents inconsistencies in input and output performance caused by context compatibility issues after switching models. Now, switching models within the same session delivers completely consistent results (excluding gaps in reasoning capabilities). For AI Agent applications, it is highly recommended to use the Primary-Backup Aggregation Mode (i.e., using multiple loadable API keys for a single model). This enhances load capacity while ensuring that the intellectual and analytical performance of the project does not experience significant drift.

2. New Client Adapter Architecture
   This version introduces the Client Adapter architecture, which enables targeted detection and optimization for clients such as Codex, Claude, and Opencode. It can deliver different Chain-of-Thought (CoT) summary block formats to different clients to achieve optimal and aesthetically pleasing rendering (provided that the thinking/reasoning mode is enabled for the model in the extra_body).

3. Cross-Protocol & Cross-Client Synergy
   Deepnow can now not only coordinate end-to-end translation and pass-through across 9 different protocols, but also detect and optimize for specific models and clients. This allows many exclusive Agents to be deployed with any model while maintaining the exact same display effects as native models (excluding reasoning capabilities).

4. "Chain-of-Thought Construction Strategy" Added to Dashboard
   A new "Chain-of-Thought Construction Strategy" has been added under Dashboard -> "Rule-based" Configuration -> "LLM Rules". This feature standardizes how text is displayed when thinking mode is enabled. While different providers use different formats (e.g., Gemini uses , DeepSeek uses reasoning, and others use think), Deepnow can now unify these variations seamlessly.

5. New SSE Watchdog Function
   Added an SSE (Server-Sent Events) watchdog function to improve stream stability and monitoring.

Minor fixes.

• Chinese Changelog:
  一个极小修正 Claude desktop 在冷启动后第一轮请求nonsse-sse 混用可能引擎dn维护的上下文工具链索引错误的id引起后续长程任务出现问题，现在Claude 可以完美工作在 Deepseek / Gemini / Kimi / Qwen / Gpt 这些已测模型家族，长程任务也不会出现任何问题.

注意：使用 Claude CLI & Desktop 时当主模型不是多模态时，次模型至少要挂载一个多模态模型. 因为Claude 的沙盒模式会进行任务截图.

另外，如果 deepnow 挂载 Deepseek 为智能体服务时，请最好在 extra body 里配置其专属 : reasoning_effort = "high" 然后点旁边的"验证"，没问题才可以保存，以提高dp在长程任务时工具链输出的稳健性.

• English Changelog:
  Minor fix for Claude Desktop regarding an issue where mixing Non-SSE and SSE in the first request turn after a cold start could cause the context toolchain index maintained by the engine to fetch the wrong ID, leading to subsequent failures in long-context tasks. Claude now works flawlessly with tested model families—including DeepSeek, Gemini, Kimi, Qwen, and GPT—ensuring long-context tasks run without any issues.

Note: When using Claude CLI & Desktop, if the primary model is non-multimodal, you must attach at least one multimodal model as a secondary model. This is because Claude's sandbox mode captures task screenshots.

btw : When mounting DeepSeek as the agent service in Deepnow, it is highly recommended to configure the exclusive parameter reasoning_effort = "high" in the extra_body. After configuring, click the adjacent "Validate Parameters" button to ensure everything passes before saving. This significantly improves Deepseek toolchain output robustness during long-context tasks.

Claude fully support !

Claude CLI & Claude Desktop 全面支持！ 至此，Deepnow 已完整支持全球已发布的所有智能体和全部端点协议。

目前 Deepnow 的协议库支持的协议有v1: Completions / Responses / Messages , Deepnow 正在维护一个极其复杂的协议中间层。

客户端协议 ↓ \ 端点协议 →	R 协议模型	C 协议模型	M 协议模型
R 协议客户端	同协议透传 (R → R)	异构协议翻译 (R → C)	异构协议翻译 (R → M)
C 协议客户端	异构协议翻译 (C → R)	同协议透传 (C → C)	异构协议翻译 (C → M)
M 协议客户端	异构协议翻译 (M → R)	异构协议翻译 (M → C)	同协议透传 (M → M)

• 同时除了 9 种组合的连接方式，每种方式还同时支持 SSE 和 Non-SSE ，可以说现阶段是唯一如此完整支持所有AI端点协议且内置RAG+Server side 上下文的算力底座。

• 意味着你永远不用关心你的AI应用使用什么样的协议连接 Deepnow 都可以实现完整支持，同时你也不用关心你手头的端点模型是否能支持你的应用，总之现在你可以用 Claude 挂载 本地模型或者 deepseek 来玩，也可以用 Codex 去挂载一个 opus 。

---

Claude CLI & Claude Desktop are now fully supported! With this release, Deepnow delivers comprehensive compatibility with all globally deployed AI agents and endpoint protocols.

Deepnow's current protocol library supports v1 standards: Completions / Responses / Messages. To seamlessly bridge these distinct architectures, Deepnow actively maintains a highly sophisticated protocol middleware layer.

• Furthermore, in addition to the nine discrete routing matrix combinations, each operational pathway concurrently supports both Server-Sent Events (SSE) and non-SSE streaming modalities. Consequently, at the current evolutionary stage, Deepnow represents the definitive high-performance compute foundation capable of providing such exhaustive compatibility across all AI endpoint protocols while natively synthesizing inline Retrieval-Augmented Generation (RAG) and stateful server-side context preservation.

• This architectural abstraction insulates downstream client applications from protocol heterogeneity; any AI agent interfacing with Deepnow receives full native support, irrespective of the transaction schema utilized. Concurrently, developers are decoupled from the compliance and compatibility constraints of upstream endpoint models. In practice, this enables platforms like Claude Desktop to seamlessly orchestrate localized weights or DeepSeek models, or conversely, allows frameworks like Codex to effortlessly mount an Anthropic Claude Opus endpoint.

Model Family / Model Adapter / More Custom Provider Added.

大语言模型配置方式更改、模型适配器增加

1. 大语言模型的配置页面调整为更加合理的方式，自定义模型端点接入可以扩展为最多5个，可以用于deepnow级联或者llamacpp / lmstudio 等更多本地推理机连接；
2. Deepnow 本身已经在多模型协商算法上做了大量的工作，包括上下文对齐、call_tool_id 对齐，chunk 分段投递，sse 级的多模型切换等，此版本又新增了一个模型家族适配器算法，目的是为完全实现对客户端(Agent)的归一化的响应能力，所以不管连接任何模型我们通过3段式维护方式来达到此设计目的，故引入adapter算法：

• 第一段,由上下文维护状态机进行分析投递和协商目标模型和客户端之间的连接，根据模型能力决定翻译还是透传；
• 第二段,由Adapter 针对模型特性优化，比如gemini 的singature 和deepseek 的reasoning_content 机制，以Model Family的形式检测和优化向模型投递请求时的精准构建；
• 第三段,由用户可自增携带模型特性的开关特性可放入 extra_body 来构建每次投递模型特有参数或开关的载荷，比如你可以在模型配置的extra_body 里为deepseek 模型新增: reasoning_effort="high" 这个参数，然后你还可以通过右侧的“验证”功能实时测试这个参数是否合法，如果验证后上游模型回复成功，则这个参数可以变为有效载荷来进行投递，以此使模型的工作更加符合预期。
• 同时,一个项目可以在任意一个完整工作流结束时随时切换模型，让项目可以被多个模型接管，以实现按需使用低、中、高能力模型的使用，以后还会增加这种自动检测的能力。你甚至还可以挂载多个单功能模型让推理和图形识别、音频分开调用不同模型，但是客户端是无感的，客户端始终认为是一个模型在工作。
• 负载模式（如聚合、混合）时不在一个delta 机制内的任务都可以被切换和负载，上下文的历史会被投递到多个模型，当多人使用deepnow 时可以让拥有多个模型的用户实现多应用、多智能体的高速响应能力。

3. 此版本以后将不再进行构架方面的调整，后续将专注能力的开发，所以可以将此版本作为里程碑版本使用；

--- English Changelog ---

1. LLM Configuration UI Overhaul & Extended Endpoint Support
   Optimized Configuration Interface: Redesigned the LLM configuration page to deliver a more intuitive and rational user experience.

Expanded Custom Endpoints: Added support for up to 5 concurrent custom model endpoints. This enables seamless cascading with Deepnow or integrations with local inference engines such as llama.cpp and LM Studio.

2. Architecture Upgrade: Model Family Adapter & Multi-Model Negotiation
   To achieve fully normalized and transparent response capabilities for downstream clients (Agents), Deepnow builds upon its existing multi-model negotiation foundation (including context alignment, call_tool_id alignment, chunk-based delivery, and SSE-level model switching) by introducing a new Model Family Adapter algorithm.

This mechanism guarantees client-agnostic interaction across heterogeneous LLMs via a 3-stage lifecycle management framework:

Stage 1: Context State Machine Negotiation
The context-aware state machine analyzes payloads, coordinates connections between the target model and the client, and dynamically decides whether to transparently pass through or translate the payload based on model capabilities.

Stage 2: Model Family Adapters
Tailored optimization layers designed for specific model architectures (e.g., handling Gemini's signature mechanism or DeepSeek's reasoning_content). This stage auto-detects and optimizes payload construction to match the precise requirements of each model family.

Stage 3: Dynamic Parameters via extra_body & Real-Time Validation
Users can now pass model-specific hyperparameters or feature toggles using the extra_body field.

Example: You can append reasoning_effort="high" for DeepSeek models within the configuration.

Validation: A newly introduced "Verify" feature allows real-time testing of these custom parameters against upstream APIs. Once verified successfully, the parameter is locked into the active payload, ensuring predictable model behavior.

Additional Orchestration Enhancements:
Dynamic Workflow Switching: Projects can now switch models instantly at the completion of any full workflow node. This enables cost-effective, on-demand routing between low, medium, and high-tier models. (Automated detection and routing capabilities will be expanded in future releases).

Multi-Model Decoupling (Client-Agnostic): You can bind multiple specialized models to handle reasoning, vision, and audio tasks separately. Downstream clients remain completely unaware of the underlying switching, interacting with the system as a single unified model.

Advanced Load Balancing & Aggregation: For non-monolithic delta mechanisms, tasks can be dynamically switched and load-balanced under Aggregated or Hybrid modes. Context history is automatically synced across multiple target models. This ensures high-throughput, low-latency responsiveness for multi-app and multi-agent scenarios when multiple users access Deepnow simultaneously.

3. Architecture Freeze & Milestone Release
   Stability Baseline: This release marks the finalization of major architectural and structural overhauls. Moving forward, development will focus exclusively on feature enhancement and capability expansion.

Milestone Status: This version is designated as a Milestone Release and is highly recommended as a stable baseline for production deployments.

Internationalization (i18n) Support

Changelog:

1. 多语言支持 增加i18n多语言支持，可以自定义挂载语言文件。默认当前内置 “英语” “简体中文” “日语” 3个语言。系统第一次运行时默认英文，可以通过 "System Management" -> "Interface" 来选择语言，选择以后自动记忆。注意：由于deepnow网关有极强的隐私熟悉，所以语言选择并不跟随浏览器而是管理员设定，也就是说语言设置是全局的而非访问Dashboard的浏览器跟随。

2. 完成 DeepNow 全链路多语言与协议提示优化，补齐模型探测、Extra Body 验证及动态错误信息翻译，模型协议探测和载荷探测更加准确和细致，可以做到完全傻瓜式配置模型。

3. 增加一个Provider Adapter，对每一个模型进行深度兼容性调优，在非透传模式下可以利用deepnow的协议翻译机能使所有模型稳定工作在智能体模式下。

4. 聚合模型并行能力优化，现在配合智能体类的工具可使推理任务和工具调用合成效率提高，但可能会多耗费少量的输入TOKEN，但不会增加输出TOKEN。智能体模式下推荐设置成主次模型而非混合模型，其它推理任务可按需设置混合模型。

5. 智能体执行时如遇到模型出现相关能力缺失、访问失败等所有错误（如：没有多模态能力）则会自动无缝滑动到次模型尝试，所有模型均尝试失败才会抛出失败信息，以此提高智能体连续工作的稳定性。

6. 流式机闭合优化，缩短负载锁，优化并发稳定性。

7. 二进制包已包含 Mac (非Intel) / Openwrt AMD64 / Linux AMD64 / Windows AMD64

注意： 没有特别说明的情况下，旧版更新皆直接覆盖二进制可执行文件即可(早于20260701之前的老版本更新701或702必须清除一次system.db/shm/wal文件，如701更新时已清除则可忽略)，但访问Dashboard后注意浏览器使用Ctrl+F5强制刷新页面，否则新版本的更新内容可能会由于Cache无法显现。

• 当前目录下没有任何deepnow 相关文件，deepnow server 则视为第一次使用。

English Changelog:

1. Internationalization (i18n) Support
   Added i18n multi-language support, allowing for custom mounting of language files. By default, three languages are built-in: English, Simplified Chinese, and Japanese. On the system's first run, the default language is English. You can select your preferred language via "System Management" -> "Interface", which will then be automatically remembered.

Note: Due to the strong privacy-oriented nature of the DeepNow gateway, the language preference does not follow individual browser settings. Instead, it is configured by the administrator—meaning the language setting is global, rather than per-browser session accessing the Dashboard.

2. End-to-End i18n & Protocol Prompt Optimizations
   Completed full-link (end-to-end) multi-language and protocol prompt optimizations for DeepNow. Added comprehensive support for model probing, Extra Body validation, and dynamic translation of error messages. Model protocol and payload probing are now significantly more precise and detailed, enabling completely foolproof, "plug-and-play" model configurations.

3. Provider Adapter Addition
   Introduced a Provider Adapter to perform deep compatibility fine-tuning for each model. In non-passthrough mode, it leverages DeepNow's protocol translation capabilities to ensure all models operate stably within Agent Mode.

4. Aggregated Model Parallelization Optimization
   Optimized parallel capabilities for aggregated models. When paired with agentic tools, the synthesis efficiency of inference tasks and tool calls is significantly improved. While this may consume a slightly higher amount of input tokens, it will not increase output tokens. In Agent Mode, it is highly recommended to configure Primary/Secondary models rather than Hybrid models; other standard inference tasks can still utilize Hybrid models as needed.

5. Seamless Agent Failover (Seamless Sliding)
   During agent execution, if a model encounters any capability deficiencies, access failures, or other errors (e.g., lack of multimodal capabilities), the system will automatically and seamlessly slide over to the secondary model to retry. A failure message will only be thrown if all configured models fail, substantially improving the continuity and stability of agent workflows.

6. Stream Mechanism Closure Optimization
   Optimized the streaming mechanism closure logic to shorten payload locks and improve concurrency stability.

7. Binary Pack Distributions
   Binary packages now include support for: Mac (Apple Silicon / Non-Intel), OpenWrt AMD64, Linux AMD64, and Windows AMD64.

Note:
Upgrading: Unless explicitly stated otherwise, updates can be applied by simply overwriting the existing binary executable. (For legacy versions older than 20260701 updating to 701 or 702, you must clear the system.db, system.db-shm, and system.db-wal files once; if you already cleared them during the 701 update, you can ignore this).

Fresh Install: If there are no DeepNow-related files in the current directory, the DeepNow server will treat this as a first-time setup/clean installation.

Cache Notice: After updating and accessing the Dashboard, please remember to use Ctrl+F5 to force-refresh your browser, otherwise the new updates may not display correctly due to caching.

Wrapper & Translator Fully re-worked.

本次改动较大，请务必删除 [system.db] 相关的3个数据库文件，否则将可能导致不兼容. (本次暂时只提供 Linux 和 Windows 版本，待下个版本重新提供全平台）

1. 增加 Extrabody ，可对模型配置时自定义 payload struct
2. 增加 针对模型可支持自定义协议探测（是否支持 Completions & Responses）
3. 增加 透传和Wrapper 翻译机优化（C->C 透传 / R->R 透传 / C->R Wrapper / R->C Wrapper)
4. 针对智能体（如Codex,Hermes） 挂载Deepseek 或者 本地 Qwen3.6 翻译机的完整 tool_calls 能力和熔断能力 ，本地模型不会因为-jinjia格式或上下文不够输出导致的tool calls loop，让本地模型能够真正意义的在codex 工作正常
5. Deepseek 在智能体下不会因为reasoning_content 引起的中断
6. Deepnow 现在保存上下文（30分钟TTL)，用于负载均衡时格式快速转换和call id的核对，并会根据模型的不同投递时构造专属上下文格式（比如deepseek）应该是所有多模型路由网关里唯一实现可以上下文无缝投递给任意模型的能力。
7. 负载均衡系统（多key聚合或混合模型）可以正确给智能体分担命中，并在多个智能体任务并行时可以切实提高效率。( 比如 codex 很多任务会同时发出2个请求线程）
8. 优化 gemini 低端模型对智能体的兼容性

[English Log]
CRITICAL: This update introduces major changes. Please make sure to delete the 3 database files associated with [system.db], otherwise incompatibility issues may occur.

1. Added Extrabody, allowing for custom payload struct configuration per model.

2. Added custom protocol probing per model to detect feature support (e.g., whether Completions & Responses are supported).

3. Optimized Transparent Passthrough and Wrapper translators (C->C Passthrough / R->R Passthrough / C->R Wrapper / R->C Wrapper).

4. Provided full tool_calls and circuit-breaking capabilities for AI agents (such as Codex and Hermes) when mounting DeepSeek or local Qwen 3.6 translators. This prevents local models from falling into tool-calling loops caused by Jinja formatting errors or insufficient context length, ensuring local models function properly within Codex.

5. Resolved an issue where DeepSeek would interrupt under AI agents due to reasoning_content.

6. Deepnow now preserves context (with a 30-minute TTL) for rapid format conversion during load balancing and call ID verification. It also constructs dedicated context formats (e.g., for DeepSeek) based on the target model. This is currently the only multi-model routing gateway capable of seamlessly delivering context to any arbitrary model.

7. The Load Balancing System (supporting multi-key aggregation or hybrid models) now correctly distributes hits for AI agents, effectively boosting efficiency during parallel agent tasks (e.g., when Codex spins up two request threads concurrently).

8. Optimized Gemini low-end models for better compatibility with AI agents.

New func added & more platformz support.

增加若干功能:
1) 算力配置->大模型API设置 里配置模型服务商时，现在model name可以通过点击下方的小图标直接获取该模型商的所有模型列表，直接选择即可（当然要api key正确)
2) 增加 Openclaw 的Responses协议支持，请使用 openclaw onbaord 设置custom api (openai 兼容模型), 可以选择支持模式为 v1/completion 或者更高级的 v1/responses ( 2者 deepnow 均直接支持 )，建议 responses 可以获得更快的连续响应
3) 新增 MacOS (ARM) 和 OpenWRT (x86_AMD64) 的二进制包，开箱即用。

New Features / Changelog
LLM API Model Fetching: When configuring a model provider under Compute Configuration -> LLM API Settings, you can now automatically fetch the entire model list from that provider simply by clicking the small icon below the model name field, allowing you to select your model directly (a valid API key is required).

Revamped OpenClaw Support: When using OpenClaw onboard settings to set up a custom API (OpenAI-compatible model), you can now choose the support mode as either v1/completion or the more advanced v1/responses (both are natively supported by Deepnow). We highly recommend using v1/responses to achieve faster consecutive responses.

New Platform Binaries: Added pre-compiled binary packages for macOS (ARM) and OpenWrt (x86_AMD64), ready to use straight out of the box.

RAG re-worked.

English Changelog : Here

Deepnow — 史上最强大的 AI 算力网关

更新日志

1. RAG 系统 & UI 全面重构

若从旧版本升级并已生成 rag.db，请删除所有 rag.* 相关文件（含 .whl 后缀文件）。

RAG 是本网关核心能力之一，支持完整知识管理：可添加、删除知识，精准删除指定文本块，也能根据场景临时启用/停用知识库。
本 RAG 基于 SQLite3 + 向量 C 实现，并发能力对标专业向量数据库；支持多网关级联，搭建无限扩容的分布式多级知识库。后续将新增重排序模块，进一步提升大模型推理速度。

RAG 核心特性

• 模型无关：兼容任意加载模式导入知识
• 实时启停知识库，支持高并发 Embedding，可搭建企业级高并发检索系统，且不影响网关原有能力
• 内置 RAG 滑动开关，统一控制网关所有推理任务是否加载本地知识

2. Responses API 适配优化
   修复图文多轮对话场景下，RAG 召回冗余历史对话的问题。完整兼容 OpenAI Responses 协议规范，可对接所有 OpenAI 兼容模型，适配各类基于该协议的客户端（聊天应用、智能体，支持/不支持 SSE 均可）。目前已完美适配 Codex、Hermes，包含 Codex 图片识别能力。后续将新增 Responses 透传、Messages 协议转换能力，以支持 Claude 智能体。

3. 多模型路由实现多模态能力
   主模型仅支持文本时，可挂载副模型补充视觉分析能力。网关自动路由客户端请求至对应模型处理，并返回统一格式结果，对外呈现为标准多模态推理端点。

4. 新增主流大模型预配置
   内置全品类主流大模型配置，开箱即用。

5. 问题修复
   针对全局思维链净化逻辑完成多项细节修复。

Deepnow 让任何人都可以随便搭建一个如 OpenRouter 那样的Token提供网关，支持所有主流模型且兼容 Codex / Hermes 这类的智能体，抹平所有模型的接入协议兼容性差异。

Deepnow enables anyone to easily build a gateway similar to OpenRouter. It supports all mainstream large language models, is fully compatible with agents like Codex and Hermes, and eliminates protocol compatibility differences across various models.

Codex 类 Agent Client 图片上传支持

多模态支持
Support img upload for VLM.

现在deepnow 如果挂载可以支持多模态的模型，诸如 Codex 类客户端均可以上传图片进行分析或其它编程
DeepNow now supports the integration of multimodal models. Consequently, clients such as Codex can upload images for analysis and other programming applications.

(请始终使用最新版)

Codex Full featurez SUPPORT in any custom model

Codex 自定义模型配置说明 (适合刚刚codex更新的最新版)https://github.com/cebernic/deepnow/blob/main/doc/codex.md

这个版本改动:

1. 更加完整的 OpenAI Responses 转译封装，使得codex工作更加稳定，不会出现多任务乱跳的问题
2. 引入mark3labs 的sdk 封装deepnow的mcp服务，后期会在后台增加mcp配置能力，继续保持deepnow在ai时代all in one网关的特色
3. sse 在completion & responses 模式下均使用了id，保持多人多轮多并发
4. codex 的支持除了mcp 以外基本全部能力均已支持，应该说现阶段可能是世界上唯一一个免费且完整支持codex 使用任何模型（含本地）跑编程智能体的超级网关，而且无论如何配置模型都不需要去修改codex 配置。你的所有应用都可以使用deepnow来连接和把玩。

Codex Custom Model Configuration Guide https://github.com/cebernic/deepnow/blob/main/doc/codex.md (Tailored for the latest Codex update)

What’s New in This Release:

1. Bulletproof OpenAI Response Translation & Wrapper
   We have implemented a more comprehensive translation and encapsulation layer for OpenAI-compatible responses. This drastically improves Codex's runtime stability and effectively mitigates the issue of multi-tasking crosstalk or unpredictable context jumping.

2. Deepnow MCP Service Powered by mark3labs SDK
   We’ve integrated the mark3labs SDK to wrap Deepnow's MCP (Model Context Protocol) capabilities. Looking ahead, we will introduce native MCP configuration directly into the backend UI, continuing to cement Deepnow's position as the ultimate All-in-One AI Gateway for the agentic era.

3. Multi-Tenant High-Concurrency SSE Fixes
   Server-Sent Events (SSE) now reliably utilize unique session identifiers across both completion and responses streaming modes. This ensures seamless multi-user, multi-turn, and highly concurrent execution without channel pollution.

4. The Ultimate Zero-Config Gateway for Agentic Coding
   With the sole exception of MCP (which is still rolling out), Codex is now fully feature-complete on Deepnow. Frankly speaking, at this stage, it stands as the world’s only free, production-grade super gateway that fully empowers Codex to run coding agents on any LLM—including local deployments. Best of all, no matter how you tweak your underlying models, you’ll never have to touch your Codex configuration files again. Deepnow handles everything under the hood, allowing you to seamlessly connect and experiment with all your upstream AI apps.

I will update the multi-lang inside for Deepnow.