dsh-claude-in
English | 简体中文
Use an existing Claude Code workspace in DeepSeek Harness without maintaining a second copy of its configuration.
dsh-claude-in loads Rules, Skills, command Hooks, and Agent definitions directly from .claude. The source files remain unchanged, making .claude the single source of truth across Claude Code and DSH.
Highlights
- No migration — Claude assets are read at runtime instead of copied into DSH-specific files.
- Project and user scopes — loads the current workspace
.claude together with ~/.claude.
- Native DSH integration — exposes Skills through the DSH skill service, injects Rules into sessions, and bridges command Hooks to DSH events.
- Explicit links stay valid — linked Skills and Rule imports continue to resolve outside the discovery roots.
- Optional DSH adaptations — project-specific overrides live under
.dsn, leaving Claude assets untouched.
Supported assets
| Claude Code asset | Sources | DSH integration |
|---|
| Rules | .claude/rules/**/*.md, ~/.claude/rules/**/*.md | Baseline and path-scoped session instructions |
| Skills | .claude/skills/**/SKILL.md, ~/.claude/skills/**/SKILL.md | Native DSH skills, read from source when invoked |
| Command Hooks | hooks.json, settings.json, settings.local.json | DSH lifecycle hooks |
| Agents | .claude/agents/**/*.md, ~/.claude/agents/**/*.md | Agent catalog and claude-agent-* skills |
Session history, MCP configuration, Claude plugins, commands, memory, and direct CLAUDE.md injection are outside this plugin's scope.
Install
Version 0.2.0 targets DeepSeek Harness 0.2.1-rc.2 and its awaited agent/created lifecycle. This is the verified compatibility baseline; later DSH versions have not been validated for this release.
GitHub source
dsh plugin --profile web add github:ShinewineW/dsh-claude-in#v0.2.0
Git-based installation builds the package through its prepare script. pnpm 10 or later may require the build to be allowed in the target profile:
# $DSH_HOME/profiles/web/pnpm-workspace.yaml
allowBuilds:
dsh-claude-in: true
After adding the entry, run the installation command again.
Prebuilt release
The v0.2.0 tarball includes prebuilt code and does not require install-time build permission:
curl -LO https://github.com/ShinewineW/dsh-claude-in/releases/download/v0.2.0/dsh-claude-in-0.2.0.tgz
dsh plugin --profile web add ./dsh-claude-in-0.2.0.tgz
How it works
Discovery has two entry points:
.claude directly under the current DSH workspace~/.claude
Parent directories and unrelated filesystem locations are not scanned. A session started from a nested directory does not inherit a .claude directory located higher in the tree.
Symbolic links within discovered Skill and Agent trees are followed. Rule @import directives may also resolve files or directories outside the discovery roots. These are explicit asset references, not additional discovery locations.
Context timing
Rules and the Agent catalog are included in the first admitted model request of new and resumed sessions. Skills are registered when the plugin loads. UserPromptSubmit Hook context joins that same request, as does SessionStart Hook context when the plugin observes the awaited agent/created lifecycle. Dynamically loading the plugin after that event cannot replay SessionStart. No preset adds a first-request delay.
Rules
Rules without paths or globs are included in the session baseline by default. alwaysApply: true provides the same behavior explicitly. Path-scoped Rules are matched when DSH read, write, or edit tools touch a file and are injected once per session after the matching tool result.
Project and user Rules are combined. Project Rules are rendered first and receive priority if the configured context budget is exhausted.
Skills
Each valid SKILL.md is registered with the DSH skill service. The original file is read when the Skill is invoked, so updates remain in .claude and no managed copy is created.
Name precedence, from highest to lowest:
- Project
.dsn/skills
- User
~/.claude/skills
- Project
.claude/skills
Hooks
The following Claude command Hook events are bridged:
SessionStartUserPromptSubmitPreToolUsePostToolUseStopSubagentStartSubagentStop
Hook groups from hooks.json, settings.json, and settings.local.json are appended in that order. Project Hooks run before user Hooks. ${CLAUDE_PROJECT_DIR} and ${CLAUDE_PLUGIN_ROOT} are resolved for the active workspace before execution.
Blocking exit code 2, permissionDecision, and additionalContext are supported. Non-command Hook types are skipped with a warning.
Agents
Agent definitions are added to a lightweight session catalog and exposed as Skills named claude-agent-<name>. Invoking one of these Skills loads the Agent definition and system prompt from its original Markdown file for use by the DSH subagent workflow.
Project Agent definitions take precedence over user definitions with the same name.
Project adaptations
DSH-specific behavior can be configured without changing Claude assets:
# .dsn/dsh-claude-in.yml
enabled: true
loadRules: true
loadSkills: true
loadAgents: true
loadHooks: true
hooks:
PreToolUse:
- from: "npx tsc --noEmit"
to: "/opt/dsh/tsc-check.sh"
Hook adaptations use exact command matching and replace only the command text. The event, matcher, timeout, and execution semantics remain unchanged.
DSH-only Skills can be placed under .dsn/skills/<name>/SKILL.md. The plugin never creates or modifies .dsn files.
Configuration
The default plugin configuration is equivalent to:
enabled: true
claudeDir: ".claude"
adaptDir: ".dsn"
loadRules: true
rulesMaxBytes: 131072
ruleMaxSourceBytes: 65536
ruleImportMaxDepth: 8
defaultAlwaysApply: true
loadSkills: true
loadAgents: true
loadHooks: true
hookTimeoutMs: 600000
stderrSummaryMaxChars: 500
claudeDir and adaptDir accept one relative directory name. Absolute paths and directory traversal are rejected.
Security
Claude assets are read-only, but command Hooks execute shell commands from project and user configuration. Project Hook execution therefore carries the same trust implications as trusting that workspace in Claude Code.
Hooks can be disabled per project:
# .dsn/dsh-claude-in.yml
loadHooks: false
Explicit Skill symlinks and Rule imports can reference content outside .claude; those referenced targets are part of the trusted configuration surface.
Known limitations
- Only command Hooks and the listed lifecycle events are supported.
- Hook outputs such as
updatedInput, systemMessage, and updatedToolOutput are not applied.
- Agent
model and tools fields are included in the bridged definition but are not enforced as native DSH runtime restrictions.
- Path-scoped Rules are injected after a matching
read, write, or edit result and cannot retroactively govern the first completed write or edit.
- File access performed inside shell commands is not visible to path-scoped Rule matching.
- Rule and Agent baselines use a short metadata cache rather than file-level watchers.
Development
npm ci
npm test
npm run build
npm pack --dry-run
The build is self-contained and does not depend on a DeepSeek Harness source checkout.
License
BSD-3-Clause. Hook execution semantics are adapted from @deepseek-ai/dsh-hooks-claude-code, also distributed under BSD-3-Clause.
