claude_to_dsh: sync Claude Code conversation history into dsh by session ID

[hanasite]

I built claude_to_dsh — a Claude Code skill that syncs your Claude Code conversation history into token-lean markdown archives that dsh can reference by session ID.

Why

Claude Code's raw session JSONL is 90%+ tool-call noise — my 100MB of history compressed to just 3.9MB of pure conversation (4%). Instead of feeding dsh the full history every time, it reads a few KB of index/summary on demand.

How it works

① checkbox confirmation (hook-enforced) → ② migrate only what's checked (--only)
→ ③ optional parallel-agent compact summaries → ④ refresh index
→ ⑤ register dsh workspaces + dsh reads proactively → ⑥ dsh references by session ID

• Hook-enforced selection: a PreToolUse hook denies any migration command without --only (note: permissionDecision: "ask" gets swallowed in acceptEdits mode, so it uses deny instead)
• Session ownership: dsh sessions belong to workspaces by cwd — headless tasks must run in the project dir with the global dsh, not pnpm dsh
• Workspace mounting: don't hand-edit workspace.json (the header index rebuilds once at startup and filters manual entries) — delete it and restart dsh web to let bootstrap rebuild everything from session cwds
• Compact step: parallel subagents write structured summaries (large sessions get head/tail + grep strategy)

Install

Paste this into any Claude Code session:

帮我安装 claude_to_dsh skill：
1. 运行 git clone https://github.com/hanasite/claude_to_dsh.git ~/.claude/skills/claude_to_dsh
2. cd ~/.claude/skills/claude_to_dsh && python install.py
3. 完成后告诉我重启 Claude Code（或 /hooks 重载）生效

Or manually: git clone https://github.com/hanasite/claude_to_dsh.git && cd claude_to_dsh && python install.py

Dependencies

• Claude Code >= 2.x, Python >= 3.10
• dsh (global): npm i -g @deepseek-ai/dsh + DEEPSEEK_API_KEY
• dsh web (optional, for workspace registration/UI)

MIT licensed. Feedback and PRs welcome!

[hanasite]

中文版介绍

claude_to_dsh —— 一个 Claude Code skill，把 Claude Code 的历史对话同步成 token 极省的 markdown 档案，供 dsh 按会话 ID 引用。

为什么做

Claude Code 的原始会话 JSONL 里 90%+ 是工具调用噪音——我的 100MB 历史压缩后只剩 3.9MB 纯对话（4%）。与其每次让 dsh 读全量历史，不如按需读几 KB 的索引/摘要。

工作流

① 勾选确认（hook 强制）→ ② 只迁移勾选的（--only）→ ③ 可选并行 Compact 摘要
→ ④ 刷新索引 → ⑤ 注册 dsh 工作区 + 主动让 dsh 读 → ⑥ dsh 按 ID 引用

• 勾选强制：PreToolUse hook 拒绝所有不带 --only 的迁移命令（注意 permissionDecision: "ask" 在 acceptEdits 权限模式下会被吞掉，所以用 deny 强制）
• 会话归属：dsh 会话按 cwd 归属工作区——headless 任务必须在项目目录跑（用全局 dsh 而非 pnpm dsh）
• 工作区挂载：不要手动改 workspace.json（header index 启动时只重建一次，手动写入会被过滤）——删掉它重启 dsh web，bootstrap 会按会话 cwd 全量重建
• Compact 摘要：并行 subagent 生成结构化摘要（大会话用首尾 + grep 策略）

安装

粘贴下面这段给任意 Claude Code 会话：

帮我安装 claude_to_dsh skill：
1. 运行 git clone https://github.com/hanasite/claude_to_dsh.git ~/.claude/skills/claude_to_dsh
2. cd ~/.claude/skills/claude_to_dsh && python install.py
3. 完成后告诉我重启 Claude Code（或 /hooks 重载）生效

或手动：git clone https://github.com/hanasite/claude_to_dsh.git && cd claude_to_dsh && python install.py

依赖

• Claude Code ≥ 2.x、Python ≥ 3.10
• dsh（全局版）：npm i -g @deepseek-ai/dsh + DEEPSEEK_API_KEY
• dsh web（可选，用于工作区注册与 UI）

MIT 协议，欢迎反馈和 PR！

[hanasite]

关于"为什么不支持一键导入"的原因分析

现状：dsh 目前不支持把 Claude Code 的对话一键导入/迁移，这不是偷懒，是存储架构层面的差异：

1. 会话存储格式不互通：Claude Code 的会话是 ~/.claude/projects/**/*.jsonl（事件流 NDJSON，user/assistant/tool 消息混排）；dsh 用自研 SessionEvent schema（SESSION_FORMAT_VERSION=0，zstd 压缩，官方明确无兼容承诺）。两者只有"JSONL"三个字母相同，事件结构完全不同，不存在现成的映射通道。

2. 会话归属绑定 cwd：dsh 会话按工作目录（cwd）归属工作区，导入历史不仅要转换数据，还要重建 workspace 关联；而 header index 只在 web 启动时重建一次，外部进程（CLI/headless）新建的会话还会被 getter 过滤——迁移要做对，得碰这些内部机制。

3. dsh 处于 developer preview：官方当前优先打 foundation（插件架构、capability seam），兼容层和导入工具不是现在的主线，未来版本可能随时 break——这也是为什么我们选择文档引用路线而不是 hack 进 ~/.dsh/sessions/。

现阶段的做法：claude_to_dsh 走"转换 → 压缩 → 索引 → 工作区 → 按 ID 引用"的文档管线，把历史变成 dsh 能消费的轻量档案（100MB → 3.9MB），不动 dsh 内部存储。

期待：希望 dsh（或 cc-switch 的会话管理）后续能快速跟进官方一键导入——比如提供 dsh import --from claude-code 之类的迁移工具，或者双方对齐一个互通的会话交换格式。到时候 claude_to_dsh 就完成了历史使命 🎉

---

Why there is no one-click import — root cause analysis

Status: dsh does not yet support one-click import/migration of Claude Code conversations. This is not laziness — it's a storage-architecture gap:

1. Incompatible session formats: Claude Code sessions are ~/.claude/projects/**/*.jsonl (event-stream NDJSON mixing user/assistant/tool messages); dsh uses its own SessionEvent schema (SESSION_FORMAT_VERSION=0, zstd-compressed, officially no compatibility promise). The two share only the letters "JSONL" — event structures are completely different, so there is no existing mapping channel.

2. Sessions are bound to cwd: dsh sessions belong to workspaces by working directory. Importing history means not just converting data but rebuilding workspace associations — and the header index is only rebuilt once at web startup, with sessions created by external processes (CLI/headless) filtered by the getter. Doing migration right means touching these internals.

3. dsh is in developer preview: the project is currently prioritizing foundation (plugin architecture, capability seams); compatibility layers and import tooling are not the current focus, and future versions may break at any time — which is exactly why we chose the document-referencing route instead of hacking into ~/.dsh/sessions/.

Current workaround: claude_to_dsh uses a "convert → compact → index → workspace → reference-by-ID" document pipeline, turning history into lightweight archives dsh can consume (100MB → 3.9MB), without touching dsh's internal storage.

Hope: we'd love to see dsh (or cc-switch's session management) quickly ship an official one-click import — e.g. a dsh import --from claude-code tool, or an aligned interoperable session exchange format. When that day comes, claude_to_dsh will have fulfilled its mission 🎉

[bobleer]

我做了个 ACP 桥：dsh-acp-for-bitfun，把 BitFun 当 dsh 的 subagent 跑（ACP v1 over stdio），dsh 里任意会话都能用 subagent_bitfun 工具委托任务。思路跟你这个 sync 有点像，都是让别的 agent 和 dsh 互通。github.com/bobleer/dsh-acp-for-bitfun

[hanasite]

好的佬，我有时间看看，今天晚上有点晚了

[wenzhuo4657]

好奇这样做的作用是什么？

[hanasite]

可以看一下我贴上的why 部分