# dsh-memory-canvas DeepSeek Harness 插件。把体积大的工具日志卸到 `refs/*.md`,上下文里只留一张 Mermaid 画布,用 `node_id` 再召回。零守护进程。 **注意:别 pip 装错 dsh。** 官方 CLI:`npx @deepseek-ai/dsh web`。PyPI 上的 `deepseek-harness-cli` 是另一个工具(`dsh doctor` / `dsh chat`)。 GitHub topic:`dsh-plugin`。English: [README.md](README.md) ## 不装 dsh 先跑 node demo.mjs 会写 `.dsh-memory/refs/{id}.md`,打印画布,按 id 召回,并断言正文一致。 ## 装进 profile 本包是纯 JS,没有 prepare / 构建步骤,从 git 安装不需要 allowBuilds。 dsh plugin --profile web add github:guobinmengxiang-rgb/dsh-memory-canvas 先不启动、只看组合后的层: dsh --profile web --dump-config 应看见 `dsh-memory-canvas` 这一层,再按平常方式启动。 官方 CLI 是 `@deepseek-ai/dsh`(今天仍是 `0.1.0-rc.6`)。 ## 工具 智能体必须主动调用这些工具。本插件**不会**自动拦截工具日志。 | 工具 | 参数 | 结果 | |---|---|---| | `memory_offload` | `title`、`body`(必填) | 短 `id`、路径、`canvas updated` | | `memory_recall` | `id`(必填) | 存下的正文(超过 8k 会截断并注明) | | `memory_canvas` | 无 | 任务画布的 Mermaid Markdown | 默认存储:`/.dsh-memory/canvas.md` 与 `/.dsh-memory/refs/{id}.md`。可用 `DSH_MEMORY_DIR` 改根目录。 画布节点标题最多 24 个字符。全文在 `refs/`。 ## 这不是什么 - 不是 [TencentDB Agent Memory](https://github.com/TencentCloud/TencentDB-Agent-Memory)(不是他们的 L0-L3 人设流水线) - 不是 OpenClaw - 不是向量数据库 - 不是 [Yiipu/dsh-agentmemory](https://github.com/Yiipu/dsh-agentmemory)(那个需要 `:3111` 上的守护进程) 灵感来自腾讯的短时符号记忆(MIT 思路,代码是我们的):refs + Mermaid 画布 + `node_id`。 ## 诚实边界 - 这是开发者预览:THERE WILL BE COMPATIBILITY-BREAKING CHANGES. - 插件在你的机器上跑。 - 本组合包的 `cordis.patch.yml` 用的是包名 `dsh-memory-canvas`,不是文件系统路径。 ## 官方文档与相关 - 第一个插件:https://deepseek-harness.github.io/deepseek-harness/en/develop/basic/ - 打包与安装:https://deepseek-harness.github.io/deepseek-harness/en/develop/basic/publish - Cordis:https://github.com/cordiverse/cordis - 姊妹插件:https://github.com/guobinmengxiang-rgb/dsh-xhs-note - 课程:https://github.com/guobinmengxiang-rgb/dsh-getting-started - 精选列表:https://github.com/guobinmengxiang-rgb/awesome-dsh ## 许可 MIT (c) 2026 国彬 张