1.3.0 - 多 Agent 运行时与本地记忆 / Multi-Agent Runtime and Local Memory

v1.3.0

中文

稳定可用的多 Agent 工作区： v1.3.0 把 Codex、OpenCode 和 Hermes 的安装、运行状态、对话过程、本地记忆与轻量模型配置收进同一套 EchoInk 体验。切换后端只改变执行引擎，不再改变用户看到的对话结构。

三后端，一套运行时

• 新增三 Agent 安装与连接面板：统一查看 Codex、OpenCode、Hermes 状态，并从缺失项直接进入安装、修复或重新检测。
• 三个后端统一展示公开推理、工具调用、处理时长和最终回答；收起时优先阅读结论，展开后再查看完整过程。
• 加强 OpenCode 会话恢复、Hermes 工具终态和中断处理，避免断流后重复执行同一个任务。

Memory V2 与知识规则

• 新增 EchoInk Memory V2：在本地提炼并保存跨会话、跨后端可复用的信息，同时保留各 Agent 自己的原生记忆。
• 知识库任务每轮都会重新读取设置中指定的 Markdown 规则文件；文件缺失、为空、越界或不可读时会在 Agent 启动前停止。
• 普通 Vault 不再依赖 AGENTS.md，默认继续使用 LLM-WIKI.md，也可以在设置中选择其他 Vault 内 Markdown。

提示词增强与轻量模型

• 提示词增强拥有独立的 Agent 后端和模型设置，不会改动普通聊天的主 Agent 模型。
• 增强模型改为可选择列表；点击“新增模型”输入模型 ID 后，会立即加入当前后端列表并选中。
• Codex 默认使用 gpt-5.6-terra、中等思考和快速响应；OpenCode 与 Hermes 只显示各自实际支持的模型、Agent 或 Profile 选项。

修复与可靠性

• 修复中断的 Memory 事务可能留下不完整状态的问题，并明确本地历史与 Agent 原生记忆的边界。
• 修复不同提示词增强后端之间残留不兼容模型、Agent 或 Profile 选择的问题。
• 修复不支持思考强度或快速档位的后端仍显示对应控件的问题。

使用方法

1. 安装 v1.3.0 并重载 Obsidian。
2. 打开 EchoInk 设置，在顶部确认至少一个 Agent 已就绪；需要时安装或重新检测其他后端。
3. 从侧栏顶部切换 Codex、OpenCode 或 Hermes，当前 EchoInk 会话不会被清空。
4. 在“提示词增强”中保留自动模型，或点击“新增模型”加入自己的模型 ID。
5. 需要跨会话本地记忆时，在设置中初始化并同步 Memory V2。

现有 Vault 文件和 EchoInk 会话无需迁移。

下载

Obsidian 官方安装会读取本 Release 下的单文件资产：main.js、manifest.json、styles.css。

English

A production-ready multi-Agent workspace: v1.3.0 brings setup, runtime status, conversation projection, local memory, and utility-model configuration for Codex, OpenCode, and Hermes into one EchoInk experience. Switching backends changes the execution engine without changing the conversation layout.

Three backends, one runtime

• Added a unified setup and connection dashboard for Codex, OpenCode, and Hermes, with direct install, repair, and recheck paths for missing requirements.
• All three backends now share the same presentation for public reasoning, tool calls, processing time, and final answers. Keep the process collapsed for answer-first reading or expand it for full detail.
• Hardened OpenCode session recovery, Hermes tool terminal states, and interrupted runs so a dropped stream does not execute the same task twice.

Memory V2 and Knowledge rules

• Added EchoInk Memory V2 for locally curated, cross-session and cross-backend recall while preserving each Agent's native memory.
• Every Knowledge run now reloads the configured Markdown rules file. Missing, empty, unreadable, out-of-vault, or invalid files stop before the Agent starts.
• Ordinary Vaults no longer depend on AGENTS.md. LLM-WIKI.md remains the default, and another Vault Markdown file can be selected in settings.

Prompt enhancement and utility models

• Prompt enhancement now has its own Agent backend and model settings without changing the main chat model.
• The enhancer model is now selected from a list. Use Add model to enter a model ID, add it to the current backend, and select it immediately.
• Codex defaults to gpt-5.6-terra, medium reasoning, and fast responses. OpenCode and Hermes show only the model, Agent, or Profile controls they actually support.

Reliability fixes

• Fixed interrupted Memory transactions leaving incomplete state, and clarified the boundary between local history and native Agent memory.
• Fixed incompatible model, Agent, or Profile selections carrying across prompt-enhancer backends.
• Fixed unsupported reasoning or fast-response controls appearing for backends that do not provide them.

How to use

1. Install v1.3.0 and reload Obsidian.
2. Open EchoInk settings and confirm that at least one Agent is ready. Install or recheck additional backends when needed.
3. Switch between Codex, OpenCode, and Hermes from the sidebar header without clearing the current EchoInk session.
4. Keep the automatic prompt-enhancer model, or use Add model to register your own model ID.
5. Initialize and sync Memory V2 in settings when you want local recall across sessions.

Existing Vault files and EchoInk sessions do not require migration.

Download

Obsidian reads these single-file Release assets directly: main.js, manifest.json, and styles.css.