dsh-skill-forge
面向 DeepSeek Harness 的多智能体协作技能锻造系统——将对话经验提炼为可验证、可复用的技能，并通过进化优化进行改进。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Epiphany-Leon/dsh-skill-forge#6f206cbb79f7cfe674372f4b3979c4aa99c5e389Skill Forge 以 dsh-skill-forge 发布，当前版本为 0.3.0。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。
🌏 中文文档 | 中文版
Multi-Agent collaborative skill forging system for DeepSeek Harness. Distills conversational experience into verifiable, traceable, reusable Agent Skills through engineering-grade quality control.
Each skill passes through seven quality gates before activation: trigger, extract, generate, verify, iterate, audit, and approve. The final output meets practical usability standards instead of surface-level correctness.
Six specialized agents work in a coordinated pipeline. The Extractor distills patterns from conversations. The Generator drafts skill documents. The Verifier runs multi-dimensional validation. The Refiner performs incremental optimization based on feedback. The Adversarial Test Generator creates challenging test cases for co-evolution. The Auditor handles security and duplication checks. Each role carries independent prompts and evaluation criteria for full transparency.
Built on the insight of SkillOpt-style textual gradient descent. When verification fails, the Refiner Agent locates specific issues from failure patterns and applies targeted rewrites instead of full regeneration. The system enforces non-regressive quality across iterations, with a configurable maximum round count (three by default).
Three evolution engines continuously improve the skill library:
Off-peak batch optimization that runs during idle time. The engine performs health checks on the skill library, batch-optimizes low-scoring skills, auto-fuses similar skills, and generates improvement suggestions. Three trigger modes are supported: manual, cron schedule, and idle detection. All changes follow strict non-regression and never auto-delete.
Complex task decomposition and multi-skill workflow composition. The orchestrator breaks tasks into subtasks (fast keyword matching or LLM-based semantic decomposition), matches relevant skills from the library, and generates execution guidance. Supports up to 5 skills per task with configurable relevance thresholds.
Security coverage spans the full skill lifecycle: generation-time, verification-time, storage-time, and usage-time safety. Auto-generated content receives default distrust. Dangerous pattern matching, duplicate detection, and manual approval form three layers of guardrails. All evolutionary modes default to human-in-the-loop approval.
Every skill carries a complete family tree: source conversation IDs, forge round records, per-round verification scores, version changelogs, usage feedback statistics, and lineage graphs (derived skills, related skills). Full auditability lets you trace where a skill came from, how it evolved, and how it performs.
Five-dimensional weighted scoring combines verification score, usage frequency, keyword match, freshness, and feedback score with token budget management. Smart injection replaces the blanket registration approach by dynamically selecting the most relevant skills per conversation turn. Reward-driven weight auto-adjustment learns from both explicit and implicit user feedback.
Deeply embedded in the DSH Web UI. A right sidebar hosts the forge queue, skill library, file browser, statistics dashboard, and settings. A top bar shows forge status with quick actions. Toast notifications push real-time progress updates with one-click approve/reject. The entire workflow from trigger to approval completes inside the conversation interface.
Skill Forge is installed as a DSH plugin via your profile's package.json.
Add the plugin to your DSH profile:
# In your DSH profile directory (e.g. ~/.dsh/profiles/web)
pnpm add dsh-skill-forge@github:Epiphany-Leon/dsh-skill-forge
Then edit your profile's package.json to enable the bundle:
{
"dsh": {
"profile": {
"bundles": [
"@deepseek-ai/dsh-base",
"@deepseek-ai/dsh-web-app",
"dsh-skill-forge"
]
}
}
}
Restart DSH — the Skill Forge panel appears in the right sidebar.
git clone https://github.com/Epiphany-Leon/dsh-skill-forge.git
cd dsh-skill-forge
pnpm install
pnpm build
# Link to your DSH profile
cd ~/.dsh/profiles/web
pnpm add dsh-skill-forge@link:/path/to/dsh-skill-forge
# Then add "dsh-skill-forge" to dsh.profile.bundles in package.json
# Start DSH Web
dsh web
# Open DSH in your browser. The Skill Forge panel appears in the right sidebar.
forge_skill tool directly.See docs/quickstart.md for a detailed walkthrough (Chinese).
Adjust settings through the DSH settings UI, or edit the profile cordis.patch.yml directly.
| Option | Type | Default | Description |
|---|---|---|---|
securityLevel | strict / normal / permissive / auto | normal | Security audit strictness |
autoTrigger | boolean | true | Enable automatic forge triggering |
triggerThreshold | number (0.1–0.95) | 0.3 | Confidence threshold for auto-trigger |
maxIterations | number (0–10) | 3 | Maximum forge iteration rounds |
verificationPassThreshold | number (0.5–1.0) | 0.9 | Verification pass threshold |
tokenBudgetRatio | number (0.05–0.3) | 0.1 | Token budget share for skill injection |
injectionMode | all / smart | all | Skill injection mode: full registration or smart dynamic injection |
injectionRelevanceThreshold | number (0–1) | 0.2 | Minimum relevance for smart injection |
enableDreaming | boolean | false | Enable off-peak batch forging |
dreamingSchedule | string | 0 3 * * 0 | Cron schedule for off-peak forging |
customDangerousPatterns | string[] | [] | Custom dangerous pattern list |
workspaceRoot | string | '' | Skill library storage root (uses first DSH workspace if empty) |
Full configuration reference: docs/configuration.md (Chinese).
Session Events → [Trigger] → [Extract] → [Generate] → [Verify] → [Audit] → [Approve] → Registry
↓ ↑
[Refine] ───┘
dsh-skill-forge/
├── src/
│ ├── index.ts # Host entry (Cordis plugin)
│ ├── types.ts # Global type definitions
│ ├── prompts/ # Agent System Prompts
│ ├── services/ # Host services (12 modules)
│ │ ├── ForgeOrchestrator.ts # Core orchestrator
│ │ ├── TriggerEngine.ts # Incremental trigger engine
│ │ ├── SkillRegistry.ts # Versioned skill registry
│ │ ├── InjectionEngine.ts # Smart injection engine
│ │ ├── SecurityAuditor.ts # Security auditor
│ │ ├── SkillForgeService.ts # Public service + HTTP routes
│ │ ├── DarwinOptimizer.ts # Darwin mode - hill climbing
│ │ ├── TaotieFusion.ts # Taotie mode - skill fusion
│ │ ├── CoEvoOrchestrator.ts # CoEvo mode - co-evolution
│ │ ├── DreamingEngine.ts # Dreaming mode - batch optimize
│ │ └── SkillOrchestrator.ts # Skill orchestration
│ ├── agents/ # LLM Agents (6 roles)
│ │ ├── ExtractorAgent.ts
│ │ ├── GeneratorAgent.ts
│ │ ├── VerifierAgent.ts
│ │ ├── RefinerAgent.ts
│ │ └── AdversarialTestGenerator.ts
│ └── client/ # React client (15+ components)
│ ├── index.tsx
│ └── components/ # UI components
├── docs/ # Documentation
│ ├── quickstart.md
│ ├── configuration.md
│ └── faq.md
├── cordis.patch.yml # DSH plugin configuration
├── tsdown.config.ts # Build configuration
├── CHANGELOG.md # Release notes
├── STATUS.md # Project progress tracking
├── CONTRIBUTING.md # Contribution guide
└── package.json
Contributions of code, documentation, and feedback are welcome.
# Clone
git clone https://github.com/Epiphany-Leon/dsh-skill-forge.git
cd dsh-skill-forge
# Install and build
pnpm install
pnpm build
# Type check
pnpm typecheck
# Install into local DSH profile for testing
dsh plugin --profile web add .
This project builds on ideas from several pioneering projects and research papers.
Full acknowledgments (in Chinese) are in the Chinese README.
MIT © Epiphany-Leon