v0.4.4

Session Atlas

• DeepSeek Harness joins Claude Code and Codex as a built-in source. Sessions under ~/.dsh/sessions are discovered by project folder, listed with the dsh tag, and open as a mirror: one node per message you actually sent, the model's answer for that turn, and every tool call paired with its result as an attachment.
• The harness writes its logs as zstd concatenated frames. The desktop shell decodes them frame by frame with no external tool. A frame still being written is skipped and read whole on the next sweep.
• Runtime-context and skill-catalog messages the harness injects ahead of each turn are recognized as template and never open a turn. Reasoning text is not shown as the answer.
• A session file that has no message yet (the harness creates one as soon as a session is opened) is not listed.
• Not in this release: the thoughtdag CLI and its MCP tools do not index DeepSeek Harness sessions yet.

---

Agent 对话地图

• DeepSeek Harness 成为内置来源，与 Claude Code、Codex 并列。~/.dsh/sessions 下的会话按项目文件夹发现，带 dsh 标签，打开即镜像：你真正发出的每条消息一个节点，配那一轮的模型回答，每次工具调用与结果配对成附件。
• Harness 的日志是 zstd 拼接帧，桌面壳逐帧解码，不依赖外部工具。正在写入的尾帧先跳过，下次扫描整帧读入。
• Harness 在每轮前注入的运行时上下文和技能目录消息被识别为模板，不会开启轮次。推理文本不会显示为回答。
• 尚无消息的会话文件（打开会话时 Harness 就会创建）不列出。
• 本版未包含：thoughtdag 命令行与 MCP 工具尚未索引 DeepSeek Harness 会话。