- 产品侧一直并存 单代理 ReAct（走 /api/agent-loop）与 Eino 多代理里的 Deep（走 /api/multi-agent*，作为多代理的基线编排）。在此基础上，本次新增的是多代理的另外两种编排：Plan-Execute（规划–执行–重规划）与 Supervisor（主代理 transfer/exit 监督子代理），三者仍由请求体 orchestration 与前端/WebShell 等入口选择，并与原有 ReAct、Deep 共用 MCP 与会话体系。
- Skills 侧从「零散提示」收拢为 Eino 官方 skill 渐进式披露：以全局 skills_dir 为根目录，由 multi_agent.eino_skills 控制是否启用中间件、skill 工具显示名以及是否挂载 本机 read_file/glob/grep/write/edit/execute（与 eino-ext local backend 共用同一套磁盘后端，便于在授权环境里按需暴露主机能力）。技能包按目录与 SKILL.md 等约定被索引，模型通过 skill 工具按 name 拉取，避免一次性把大量技能全文塞进系统提示；多代理子代理还可通过 bind_role / 角色 skills 列表在指令里提示优先加载的技能名。另：plantask 任务板若开启，持久化路径默认挂在 skills_dir 下的子目录（与技能包目录解耦、仅共用根路径与后端），属于工程上的目录复用而非「技能内容」本身。
- The product has long offered single-agent ReAct (/api/agent-loop) alongside Eino multi-agent Deep (/api/multi-agent*) as the baseline multi-agent orchestration. What’s new are the two additional multi-agent modes—Plan-Execute (plan → execute → replan) and Supervisor (orchestrator with transfer/exit over sub-agents)—still selected via orchestration in the API and UI while sharing the same MCP bridge and session model with the existing ReAct and Deep paths.
- Skills were refactored around Eino’s official skill progressive disclosure: a repo-wide skills_dir backs the middleware, while multi_agent.eino_skills toggles the stack, the on-the-wire tool name for skill, and optional host filesystem/shell tools (sharing the same eino-ext local backend as discovery, so file/exec capabilities are explicit and configurable). Skill packs are laid out under that tree (e.g. SKILL.md conventions) and loaded on demand via the skill tool by name, instead of inflating the system prompt with full skill text; sub-agents can still be n toward named skills through bind_role / role-level skill lists in instructions. Separately, if plantask is enabled, task JSON is persisted under a subdirectory of skills_dir—that’s path/backend reuse, not merging task state with skill documents.