---

description: "Claude Code compatibility for the DeepSeek Harness: discover .claude/skills and load CLAUDE.md rules." kind: "package-reference"

@zhang-guo-wen/dsh-claude-compat

English | 中文

Package reference. For what this plugin does and screenshots, see the repository README.

Summary

Claude Code keeps skills as /.claude/skills//SKILL.md, rules in CLAUDE.md, and scoped rules as markdown files under .claude/rules/**. This package makes the harness read all three. It registers one more skill provider on the shared registry (dsh-skill) so .claude/skills appears in the session catalog next to DSH's own roots; it folds the project .claude/CLAUDE.md and the user-global ~/.claude/CLAUDE.md into the first request as their own instructions-form context; and it folds .claude/rules/** — a paths:-scoped rule when a matching file is read, otherwise at the first request.

Table of Contents

• Use this package
• Understand the implementation
• Model Experience
• Known Limitations and Deferred Work
• Dev Note

---

Use this package

Mount the plugin alongside the skill registry; it requires ctx.skills. A profile patch or a preset composition inserts it like any other plugin.

- name: '@deepseek-ai/dsh-skill'
- name: '@deepseek-ai/dsh-claude-compat'

Field	Default	Meaning
providerName	claude-code	Unique provider name registered on ctx.skills
claudeHome	$CLAUDE_HOME or ~/.claude	Claude Code home; scanned for skills and CLAUDE.md
projectRootMarkers	['.git']	Directory entries that identify the project root
includeProjectRoot	true	Scan the project .claude/skills root
includeGlobalRoot	true	Scan the user ~/.claude/skills root
includeProjectRule	true	Load the project .claude/CLAUDE.md rule
includeGlobalRule	true	Load the global ~/.claude/CLAUDE.md rule
includeProjectRules	true	Fold the project .claude/rules/** tree
includeGlobalRules	true	Fold the user ~/.claude/rules/** tree
maxRuleSourceBytes	1048576	Maximum UTF-8 bytes read from one rule file
maxRuleRenderBytes	262144	Maximum UTF-8 bytes rendered in one rules batch

Skills

Rank	Source	Path
250	project-claude	`
/.claude/skills`
550	user-claude	~/.claude/skills

The project root is the nearest ancestor containing .git; without one the current cwd is used. Candidates carry a lower rank than DSH's own project roots, so a same-name DSH skill wins a duplicate. A skill is /.claude/skills//SKILL.md (or a flat .md) with YAML frontmatter: required name and description, plus optional whenToUse, metadata, disable-model-invocation, and user-invocable.

Rules

The project .claude/CLAUDE.md and the global ~/.claude/CLAUDE.md are read and folded into the first request as a user message under the claude-code message-source kind. The harness's own agent-instructions loads only same-directory names (AGENTS.md, CLAUDE.md); this contributor adds the two Claude Code locations without touching that loader. A missing or unreadable rule file is not fatal.

Scoped Rules

Rules under .claude/rules/** (project) and ~/.claude/rules/** (user) are folded under the claude-rule message-source kind. A rule whose YAML frontmatter carries a paths: glob list is path-scoped: it folds into the request that follows a read of a file matching one of those globs. A rule without paths (or with an empty list) is always-on and folds into the first request like CLAUDE.md. Globs are matched against the project-root-relative path; the path-scoped trigger uses the harness read tool, and any matching rule is folded at most once per session.

---

Understand the implementation

Implementation internals — click to expand

This section explains how discovery is organized; the observable behavior is fully covered in Use this package.

Design concept

The skill provider follows the skill-filesystem model: discovery parses frontmatter into catalog entries, while every load re-reads the file, so body edits need no cache invalidation. The instruction contributor reuses the agent/pre-step waterfall the way agent-instructions does, folding its instructions after the last admitted user message. Both use ctx.fs when a filesystem service is present, falling back to abortable Node I/O.

Source map

File	Role
src/index.ts	Plugin entry: registers the provider and the instruction/rule listeners
src/provider.ts	Skill provider: root resolution, discovery, and load
src/parse.ts	SKILL.md frontmatter parsing
src/frontmatter.ts	Shared YAML frontmatter parsing used by skills and rules
src/instructions.ts	Claude Code rule discovery and pre-step injection
src/rules.ts	.claude/rules/** discovery, paths: matching, and read-triggered folding
—	No runtime invariant companion is published; this package exposes no independent event sequence or mutable data relation beyond the registry and message-source contracts.

---

Model Experience

Skills catalog

What the model sees

The dsh-tool-skill tool renders this provider's invocable names and capped descriptions into the session skill catalog, source-tagged project-claude (rank 250) and user-claude (rank 550), so the model can invoke a skill by name next to every other catalog entry.

Token effect

Each skill contributes one capped description to the catalog per render. A skill's body loads only when the model invokes that skill, adding its content for that call.

KV Cache effect

The catalog is part of the tool surface and is stable across requests; a skill body loads on demand and does not change the reusable request prefix.

Rule instructions

What the model sees

The Claude Code rules and scoped rules reach the model as instructions-form user messages under the claude-code and claude-rule source kinds. Always-on rules (.claude/CLAUDE.md, ~/.claude/CLAUDE.md, and .claude/rules/** without paths) fold into the first request; a path-scoped rule with a paths: glob folds into the request that follows a read of a matching file.

Rule instruction template

Instructions from: .claude/rules/api.md

Token effect

Each folded rule is one retained user-role message bounded by maxRuleRenderBytes; a rule file larger than maxRuleSourceBytes is skipped. Always-on rules inject once per session, and a scoped rule injects once when activated.

KV Cache effect

Append-only; folded rules follow the reusable request prefix and do not invalidate existing KV-cache entries.

Known Limitations and Deferred Work

• No skill watcher yet — the provider discovers on list(); dynamic add/rename/delete of .claude/skills entries is currently picked up only when discovery runs again (for example after another provider invalidates the catalog).
• Claude/Codex rules are not reconciled after edits — unlike agent-instructions, .claude/CLAUDE.md and .codex/AGENTS.md are folded once at the first request; later edits are not re-read mid-session. Scoped rules follow the same once-per-session rule.
• Path-scoped rules trigger on reads only — matching Claude Code, a scoped rule folds when the read tool touches a matching file; write/edit do not trigger it, so a rule authoring a file it scopes may not activate.
• User-level path rules are folded on the same read trigger — the reference implementation silently ignores ~/.claude/rules path rules; this package folds both project and user trees (each at most once).
• Project scope is the nearest .git ancestor — workspaces without that marker fall back to the supplied cwd.
• Malformed entries disappear — a .claude/skills file without valid frontmatter is skipped rather than surfaced in the catalog.

Dev Note

Working context for maintainers — click to expand

None.