@guowenzhang/dsh-claude-compat
DeepSeek Harness 的 Claude Code 兼容性：技能发现、记忆文件加载、作用域规则和设置开关
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-claude-compat@0.2.1Claude Compat 以 @guowenzhang/dsh-claude-compat 发布，当前版本为 0.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 中文
DeepSeek Harness (dsh) is the open-source agent harness from DeepSeek AI, where nearly every capability is a plugin on Cordis. It is in developer preview and iterating fast, so expect compatibility-breaking changes (docs, 0.1.7-alpha.*); this plugin is a standalone third-party package that resolves @deepseek-ai/* from the running host.
Skills, memory files and path-scoped rules kept in the Claude Code .claude/ layout are invisible to a DSH session; this plugin loads all three and gives each one a switch in 设置 → Claude 兼容.
Screenshots pending — this plugin has no captured UI yet.
npx @deepseek-ai/dsh plugin --profile web add @guowenzhang/dsh-claude-compat
From the npm registry: https://www.npmjs.com/package/@guowenzhang/dsh-claude-compat — restart the host afterwards; local checkouts, git sources and troubleshooting are in AGENTS.md.
The plugin adds one section to the settings page: 设置 → Claude 兼容. It holds three switches — 加载技能, 加载记忆, and 加载规则 — and each row lists what that switch loads and when, so the page states the compatibility surface itself instead of pointing at a README. All three are on by default, and a flipped switch is a live setting: it reaches the running plugin without restarting the host.
加载技能 adds two roots to the session skill catalog: the project's <project root>/.claude/skills/** and the user's ~/.claude/skills/**. The project root is the nearest ancestor containing .git; without one, the working directory is used instead. A skill is <root>/.claude/skills/<name>/SKILL.md, or a flat <name>.md, carrying name and description frontmatter. A same-named DSH skill wins a duplicate. Discovery happens when the catalog is listed, so an added, renamed, or deleted skill is picked up on the next discovery rather than immediately.
加载记忆 folds the files Claude Code loads, broad to specific, into the request:
| Memory file | Where it comes from | When it folds |
|---|---|---|
| Managed policy | The platform's managed policy CLAUDE.md | Session start |
| User | ~/.claude/CLAUDE.md | Session start |
| Project | CLAUDE.md, .claude/CLAUDE.md, and CLAUDE.local.md in every directory from the project root down to the working directory | Session start |
| Nested | The same three names for a directory under the working directory | The request after the read tool touches a file under it |
| Auto memory | ~/.claude/projects/<repository>/memory/MEMORY.md | Session start; only the index, capped at its first 200 lines or 25 KB |
@path imports expand in place, relative to the importing file, up to four hops; a token that names no readable file stays literal, which is what leaves an @mention or an email address alone. CLAUDE.md and CLAUDE.local.md are the two names this plugin takes over: their sections are removed from the host's own workspace-instruction messages, so those files load here, under Claude Code's rules, exactly once. AGENTS.md is untouched and keeps loading from the host.
加载规则 folds .claude/rules/** for the project and ~/.claude/rules/** for the user. A rule with no paths: in its frontmatter is always-on and folds at session start, like CLAUDE.md. A rule with a paths: glob list is path-scoped: it folds into the request that follows a read of a file matching one of its globs, matched against the project-root-relative path, and at most once per session. Reading is the only trigger — write and edit never activate a scoped rule.
| Setting | Default |
|---|---|
| 加载技能 | on — .claude/skills/** and ~/.claude/skills/** are in the catalog |
| 加载记忆 | on — the memory files and the auto-memory index fold at session start |
| 加载规则 | on — always-on rules fold at session start, a paths: rule after a matching read |
.claude/skills is discovered when the catalog is listed; an add, rename, or delete is picked up on the next discovery.CLAUDE.md and no AGENTS.md keeps an intro-only reminder. The host loader's message is kept — minus the sections this plugin owns — so it stays a visible baseline and is not recomposed on every step, which leaves its one-line "the following workspace instructions may be relevant" intro with nothing after it.read only. A nested CLAUDE.md folds when the read tool touches a file in its directory; write and edit do not trigger it.read only. write and edit do not activate them.MEMORY.md and its topic files reach the model as context, but the harness does not append new memories to them the way Claude Code does.settings.json is not consulted. autoMemoryDirectory there is ignored; set the plugin's own autoMemoryDirectory field instead.AGENTS.md is not the loader's fallback. Claude Code's default reads AGENTS.md only when no CLAUDE.md or CLAUDE.local.md exists in the working directory or above it. Here the host's workspace-instruction loader keeps reading AGENTS.md on its own terms, alongside the CLAUDE.md this plugin loads.Apache License 2.0 — see LICENSE. This product includes MIT-licensed portions derived from DeepSeek Harness; see NOTICE. Not affiliated with or endorsed by Claude Code or its owners.