English | 中文
DeepSeek Harness Best Practice
A plugin-based port of the claude-code-best-practice methodology to DeepSeek Harness (DSH) — an AI agent runtime built around Cordis plugin composition.
Key Takeaway
Nearly every concept from Claude Code best practices maps onto DSH without modifying the DSH core. The implementation spans two layers:
- Static composition: agent preset (
cordis.yml) — Skills, Subagents, permissions, prompt injection
- Dynamic Cordis plugins: Hooks (event subscriptions), Status line / audio feedback (Client Slot UI)
See reports/claude-code-to-dsh-mapping.md for the full item-by-item mapping.
Repository Layout
Mapping at a Glance
| Claude Code |
DSH |
Approach |
| Skills / Commands |
Native skill catalog |
⭐ Works out of the box |
| Subagents |
subagent / workflow tools |
⭐⭐ Light adaptation |
| Hooks (20+ event points) |
Cordis Events + dynamic plugins |
⭐⭐⭐ See plugins/notify-hooks |
| Memory (CLAUDE.md / rules) |
Hindsight + CLAUDE.md injection |
⭐ Better solution already available |
| Ralph Wiggum loop |
ralph / goal tools |
⭐ Built in |
| Scheduled tasks |
dsh-task-board plugin |
⭐ Built in |
| Command→Agent→Skill orchestration |
workflow tool (code-level DAG) |
⭐⭐ Paradigm upgrade |
References
- Methodology source: shanraisshan/claude-code-best-practice (the local
framework/ directory is a reference copy of it and is not part of this repository)
- DSH plugin development: dynamic Cordis Plugin (Host events + Client Slot + Package-private RPC)
License
MIT