dsh-movein

中文

Listed in awesome-dsh-plugin · awesome-deepseek-harness · cited by dsh-handbook

Move your whole Claude Code setup into DeepSeek Harness (DSH) with one command. Skills, MCP servers, hooks and instructions arrive together, so DSH feels like home from the first prompt.

从 Claude Code 拎包入住 DSH。一条命令搬完技能、MCP、hooks 和全局指令。

dsh plugin --profile web add dsh-movein

Restart dsh web, then say "move my Claude Code setup over". The plugin exposes a movein_from_claude_code tool (dry run by default, apply=true to move).

Or run it standalone, no DSH plugin needed:

npx dsh-movein            # dry run, shows the moving estimate
npx dsh-movein --apply    # actually move in
npx dsh-movein doctor     # verify the move afterwards, any time

📦 dsh-movein · Claude Code -> DeepSeek Harness moving estimate · 拎包入住

  → CLAUDE.md (global) .............. link ~/.dsh/AGENTS.md -> ~/.claude/CLAUDE.md
  → skill computer-use .............. link -> ~/.dsh/skills/computer-use
  → skill design-taste-frontend ..... link -> ~/.dsh/skills/design-taste-frontend
  → MCP servers + hooks ............. write generated block into ~/.dsh/cordis.patch.yml

  found: 4 skills · 2 commands · 3 MCP servers · 1 hook configs · 5 subagents · 878 sessions

The full Claude Code → DSH compatibility matrix documents what every asset type does in DSH, measured against the source, including the parts no tool can automate. 完整的实测兼容性对照表见 docs/compat.md。

What moves

Asset	How
Project CLAUDE.md	Nothing to do. DSH reads it natively
Global ~/.claude/CLAUDE.md	Linked to ~/.dsh/AGENTS.md
Skills (.claude/skills)	Symlinked into DSH skill roots. The SKILL.md format is compatible as is
MCP servers (.mcp.json)	Converted to dsh-mcp-client rows in ~/.dsh/cordis.patch.yml. Tool names (mcp__server__tool) stay identical
Hooks (settings.json)	Wired through the first party dsh-hooks-claude-code bridge, your hooks config runs unchanged
Subagents (.claude/agents)	Converted to DSH skills, the system prompt becomes the skill body
Slash commands (.claude/commands)	Converted to user-invocable DSH skills, you keep typing /name
Permission rules (settings.json)	deny and ask rules enforced at DSH's tools/pre-execute gate via the companion dsh-movein-permissions plugin. allow rules stay with the DSH preset. Prefer dsh-permission-rules? --emit-rules prints your rules in its YAML shape

Skills are symlinked by default so edits keep flowing both ways. Pass --copy for real copies.

${VAR} values in MCP env become process.env references, secrets are never inlined into the generated YAML.

Reverse moving day (dual boot)

Skills you build inside DSH can come back:

npx dsh-movein --reverse            # what would come back
npx dsh-movein --reverse --apply    # bring it back

DSH-born skills land in .claude/skills (symlinked, so both sides stay current), a DSH-born AGENTS.md links to CLAUDE.md when Claude Code has none. Assets that originally moved in from Claude Code are recognized and skipped, they never left. Use both tools, keep one setup. 在 DSH 里长出来的技能可以搬回 Claude Code，双栖不二选。

After the move

npx dsh-movein doctor

A post-move health check, run it any time. It verifies that every recorded destination still exists, that no moved skill has the frontmatter shape DSH's YAML parser silently drops (#1401), that every package referenced in cordis.patch.yml still resolves (an unresolvable row is a fatal boot in DSH, not a warning), and that hook matchers will not miss DSH's lowercase tool names (#582).

If an apply left you worse off, npx dsh-movein restore puts back the newest cordis.patch.yml backup. A backup is taken automatically before every write.

Safe to run

• Dry run is the default. Nothing is written without --apply, and the estimate shows every planned change first.
• Backup before write. cordis.patch.yml is copied to ~/.dsh/movein-backups/ before each apply, restore brings it back.
• Zero dependencies, small enough to read. Plain Node, no install scripts, you can audit the whole thing before running it.
• Every move is on the record. ~/.dsh/movein-manifest.json logs source and destination per applied move, so "where did this skill come from" stays answerable.
• Secrets stay out of generated config. ${VAR} values become process.env references, and the dry run warns if a plaintext secret-looking value (API key shapes) would be copied into ~/.dsh.

What does not move

• Sessions. Out of scope here. For conversation history see dsh-chat-import.
• DSH-born MCP and hook rows (on --reverse). Hand-written cordis patches are not machine-reversible, the report says so instead of guessing.

Roadmap

Next	What
Reverse moving day	Shipped in v0.4.0, --reverse brings DSH-born skills and instructions back
doctor	Shipped in v0.5.0, post-move health check plus restore from automatic backups
More origins	Codex and Gemini CLI setups moving into DSH the same way
--watch	On hold. dsh-chat-import shipped a sync panel that covers much of this, details on the issue

A thumbs up on an issue is a vote for what ships next. Releases land fast here, four shipped in the first two days. 计划中：反向搬家（DSH 配置导回 Claude Code，双栖不二选）、--watch 双向同步、支持从 Codex/Gemini CLI 搬入。

See also

• dsh-handbook chapter 1 has a good "dsh vs Claude Code" comparison, and its ecosystem chapter lists dsh-movein for the migration path.

Verify

dsh --profile web --dump-config | grep -E "mcp-|cc-hooks"

Then restart dsh web. Counted, scanned and moved on this exact layout against dsh 0.1.0-rc.6.

Notes

• Migration diff report. Permission rules are never converted silently. The report says how many deny and ask rules are enforced, lists every rule that has no DSH-side tool to map to, and notes that * patterns match as a superset. Fail closed beats fail open.
• Hook environment scan. The dry run scans hook commands for env vars DSH will not provide ($CLAUDE_PROJECT_DIR and $CLAUDE_PLUGIN_ROOT are substituted by the bridge, anything else unknown gets a warning) so hooks do not silently break on first fire.
• Silent-drop warning. Skills whose description contains an unquoted ": " vanish from DSH's catalog without an error (#1401, full writeup in docs/skill-vanish.md). The dry run and doctor both flag that shape before it bites.
• Re-running is safe. Existing links are skipped and the generated YAML block is replaced in place, your own rows are preserved.
• Needed packages (dsh-hooks-claude-code, dsh-hook-protocol) are installed into the profile automatically, pinned to your dsh version.
• DSH is a developer preview. When plugin contracts change we patch fast.

---

中文说明

dsh-movein 把你的 Claude Code 配置一次性搬进 DeepSeek Harness。

dsh plugin --profile web add dsh-movein

重启 dsh web 后说"帮我把 Claude Code 的配置搬过来"即可（movein_from_claude_code 工具，默认预演）。也可以独立运行：

npx dsh-movein            # 预演，先看搬家清单
npx dsh-movein --apply    # 正式入住
npx dsh-movein doctor     # 搬完随时体检

权限规则迁移会输出差异报告（几条原样生效、几条无法映射都列出来），hooks 里引用的未知环境变量在预演阶段就会警告，每次搬家都记录 manifest 便于溯源。

搬什么

• 项目 CLAUDE.md 不用搬，DSH 原生就读
• 全局 ~/.claude/CLAUDE.md 链接为 ~/.dsh/AGENTS.md
• 技能目录符号链接进 DSH 技能根，SKILL.md 格式直接兼容
• 斜杠命令（.claude/commands）转换为用户可调用的 DSH 技能，/name 照常用
• .mcp.json 自动转换为 dsh-mcp-client 配置行，工具名 mcp__server__tool 完全一致
• hooks 通过官方 dsh-hooks-claude-code 桥原样运行
• 子代理（.claude/agents）自动转换为 DSH 技能
• 权限规则的 deny 与 ask 通过配套插件 dsh-movein-permissions 在 tools/pre-execute 强制执行，想用 dsh-permission-rules 的话 --emit-rules 可导出其 YAML 格式

搬完之后。doctor 检查搬过去的资产是否健在、技能会不会被 DSH 的 YAML 解析器静默丢弃（#1401）、cordis.patch.yml 引用的包是否可解析（解析不到会导致 dsh 无法启动）。每次写入前自动备份，npx dsh-movein restore 一键回滚。

会话历史不在范围内，请配合 dsh-chat-import 使用。

默认预演，不加 --apply 不写任何文件。重复执行安全，生成块原地替换，你自己的配置行不受影响。密钥永不写入生成的 YAML，预演阶段还会对疑似明文密钥告警。

Star history

Related

After moving in, see where your tokens go: dsh-lean audits a session's cache split per request and trims the DSH prompt prefix 53%. npx dsh-lean audit, nothing installed.

License

MIT