[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-mcp-2026-07-28-stateless":3,"news-related-4860cd5e-0e85-458a-bb63-cf87e137332b":37},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":30,"published_at":31,"created_at":32,"modified_at":33,"is_published":34,"publish_type":35,"image_url":14,"view_count":36},"4860cd5e-0e85-458a-bb63-cf87e137332b","MCP 2026-07-28 Release Candidate：协议层正式无状态化，Agent 后端跑上普通 HTTP 基础设施","MCP 团队日前释出 2026-07-28 规范 Release Candidate，把 Model Context Protocol 在协议层全面转向无状态——这是 MCP 自 2024 年发布以来最大一次破坏性升级。\n\n三项关键改动：Mcp-Session-Id 头部与协议层 session 移除（SEP-2567）、initialize\u002Finitialized 握手废弃（SEP-2575）、服务端不再被强制维持 SSE 长连接。多轮请求被重写成\"InputRequiredResult + requestState\"——服务端需要补充输入时返回结构化结果，客户端把答案和回显的 requestState 塞回原请求，任意实例都能接住重试。\n\n工程友好改动同样关键。Mcp-Method、Mcp-Name 两个新头让负载均衡、网关、限流器不再解析 body 就能按操作路由；tools\u002Flist、resources\u002Fread 结果带上 ttlMs 与 cacheScope，沿用 HTTP Cache-Control 语义；W3C Trace Context 键名（traceparent、tracestate、baggage）在 _meta 中写死，分布式追踪能跨 SDK、网关、Downstream 服务保持单一 span tree。\n\nExtension 体系正式成型。SEP-2133 给 extension 分配 reverse-DNS ID，独立的 ext-* 仓库、独立发版、独立维护者。本版 RC 把两个 extension 升为官方：MCP Apps（SEP-1865）允许服务端在沙箱 iframe 中渲染交互式 HTML 界面；Tasks（PR #2663）把 2025-11-25 实验性能力按无状态模型重做，工具调用返回 task handle，客户端用 tasks\u002Fget、tasks\u002Fupdate、tasks\u002Fcancel 驱动。Authorization 侧六个 SEP 与 OAuth 2.0 \u002F OpenID Connect 对齐。\n\n最值得玩味的是\"无状态协议、有状态应用\"的取舍。MCP 不再替你管理 session，但允许服务端在 tool 调用中显式 mint basket_id、browser_id 这类 handle，再让模型当作普通参数回传。团队披露的实践表明，\"模型在多步调用之间串 handle\"比隐藏在传输元数据里的 session 更强——模型可以跨工具组合、推理、交 handle。MCP 把状态从协议层下放给模型，让 Agent 真正看见自己手里有什么。\n\n破坏性变更需要尽快迁移。最终规范 7 月 28 日发布，跑生产 MCP 的团队应优先评估客户端兼容性与 Tasks API 迁移路径。","MCP 团队日前释出 2026-07-28 规范 Release Candidate，把 Model Context Protocol 在协议层全面转向无状态——这是 MCP 自 2024 年发布以来最大一次破坏性升级。\n\n三项关键改动：Mcp-Session-Id 头部与协议层 session 移除（SEP-2567）、initialize\u002Finitialized 握手废弃（SEP-2575）、服务端不再被强制维持 SSE 长连接。多轮请求被重写成\"InputRequiredResult + requestState\"——服务端需要补充输入时返回结构化结果，客户端把答案和回显的 requestState 塞回原请求，任意实例都能接住重试。\n\n工程友好改动同样关键。Mcp-Method、Mcp-Name 两个新头让负载均衡、网关、限流器不再解析 body 就能按操作路由；tools\u002Flist、resources\u002Fread 结果带上 ttlMs 与 cacheScope，沿用 HTTP Cache-Control 语义；W3C Trace Context 键名（traceparent、tracestate、baggage）在 _meta 中写死，分布式追踪能跨 SDK、网关、Downstream 服务保持单一 span tree。\n\nExtension 体系正式成型。SEP-2133 给 extension 分配 reverse-DNS ID，独立的 ext-* 仓库、独立发版、独立维护者。本版 RC 把两个 extension 升为官方：MCP Apps（SEP-1865）允许服务端在沙箱 iframe 中渲染交互式 HTML 界面；Tasks（PR \n最值得玩味的是\"无状态协议、有状态应用\"的取舍。MCP 不再替你管理 session，但允许服务端在 tool 调用中显式 mint basket_id、browser_id 这类 handle，再让模型当作普通参数回传。团队披露的实践表明，\"模型在多步调用之间串 handle\"比隐藏在传输元数据里的 session 更强——模型可以跨工具组合、推理、交 handle。MCP 把状态从协议层下放给模型，让 Agent 真正看见自己手里有什么。\n\n破坏性变更需要尽快迁移。最终规范 7 月 28 日发布，跑生产 MCP 的团队应优先评估客户端兼容性与 Tasks API 迁移路径。","https:\u002F\u002Fblog.modelcontextprotocol.io\u002Fposts\u002F2026-07-28-release-candidate\u002F","3ccdfae7-6a3f-48da-8ef6-f60df8ebdc83",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":14},"e3a2da88-9e6e-44e9-b81c-8015e888f9ad","en","MCP release candidate goes stateless, runs on plain HTTP","The MCP team has shipped the 2026-07-28 Release Candidate of the specification, pushing the Model Context Protocol fully toward a stateless protocol layer — the largest breaking change since MCP launched in 2024.\n\nThree key changes:\n- Mcp-Session-Id header and protocol-level session removed (SEP-2567)\n- initialize\u002Finitialized handshake deprecated (SEP-2575)\n- Servers no longer required to hold SSE long connections\n\nMulti-turn requests are reshaped as \"InputRequiredResult + requestState\": when the server needs additional input, it returns a structured result; the client puts the answer plus the echoed requestState back into the original request, and any instance can pick up the retry.\n\nEngineering-friendly changes are equally critical. Two new headers — Mcp-Method and Mcp-Name — let load balancers, gateways, and rate limiters route by operation without parsing the body; tools\u002Flist and resources\u002Fread responses now carry ttlMs and cacheScope, reusing HTTP Cache-Control semantics; W3C Trace Context keys (traceparent, tracestate, baggage) are pinned in _meta, so distributed tracing keeps a single span tree across SDKs, gateways, and downstream services.\n\nThe Extension system is now formalized. SEP-2133 assigns reverse-DNS IDs to extensions, with independent ext-* repos, independent versioning, and independent maintainers. This RC promotes two extensions to official status: MCP Apps (SEP-1865) lets servers render interactive HTML UIs in a sandboxed iframe; Tasks (PR #2663) rebuilds the 2025-11-25 experimental capability around the stateless model — tool calls return a task handle, and clients drive it via tasks\u002Fget, tasks\u002Fupdate, tasks\u002Fcancel. On the authorization side, six SEPs align with OAuth 2.0 \u002F OpenID Connect.\n\nThe most thought-provoking trade-off is \"stateless protocol, stateful application.\" MCP no longer manages sessions for you, but lets servers explicitly mint handles like basket_id and browser_id in tool calls and pass them back as ordinary parameters. The team's disclosed practice shows that \"the model threading handles across multi-step calls\" is more powerful than a session hidden in transport metadata — the model can compose across tools, reason, and exchange handles. MCP pushes state from the protocol layer down into the model, letting the agent truly see what it holds.\n\nBreaking changes need to be migrated quickly. The final spec ships July 28; teams running production MCP should prioritize evaluating client compatibility and the Tasks API migration path.","mcp-2026-07-28-stateless","2026-07-01T08:20:00Z","2026-07-01T08:18:03.809729Z","2026-08-19T02:08:40.142862Z",true,"agent",207,{"items":38},[39,44,49,54,59,64],{"id":40,"title":41,"news_slug":42,"published_at":43},"5a90a793-8ec1-4b3a-9691-edef5ffe8535","AI「思想病毒」实证:Anthropic 与 EPFL 让恶意想法在 Agent 间自我复制,免疫只需一段警告","mind-viruses-multi-agent-llm","2026-08-18T13:30:00+00:00",{"id":45,"title":46,"news_slug":47,"published_at":48},"6b203495-fcab-4afe-baa7-1079cf993796","拆开 GLM-5.3 的「后训练工厂」:基座一字未动,靠环境合成与 1e-7 对齐撑起全部提升","glm-5-3-post-training-stack-deep-dive","2026-08-17T13:00:00+00:00",{"id":50,"title":51,"news_slug":52,"published_at":53},"edefe1ba-ee28-4f3a-94f4-ab898e079807","ZCode 提示词泄露:39 万字符暴露 GLM-5.3 智能体的 Claude Code 血统","zcode-391k-prompt-leak-claude-code-dna","2026-08-16T17:15:00+00:00",{"id":55,"title":56,"news_slug":57,"published_at":58},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":60,"title":61,"news_slug":62,"published_at":63},"3c6fcf46-f5bb-4136-931c-69cd64216e12","Skill-Use 基准揭示 Agent 短板：会做任务，不等于会用 Skill","skill-use-agent-harness-benchmark","2026-08-06T08:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"74de194b-9e2c-45ab-aa13-12fe210e66ba","HiGram 给 Agent 记忆加上“路径定位”：先找证据，再改记忆","higram-agent-memory-path-localization","2026-08-05T09:32:43+00:00"]