dsh-plugin-custom-provider-enhancer

DSH Custom Provider Enhancer

English | 简体中文

DSH Custom Provider Enhancer is an out-of-the-box DeepSeek Harness (DSH) / Cordis plugin designed for Custom Model Providers (e.g. OneAPI, NewAPI, OpenRouter, vLLM, Ollama, and OpenAI-compatible API gateways).

When discovering or saving models in the Web GUI, it automatically queries a 3000+ model database to populate Context Window (contextWindow), Max Output Tokens (maxTokens), Multimodal Vision (input: [text, image]), and Deep Thinking / Reasoning (reasoningEfforts).

---

📋 Table of Contents

• Overview
• Compatibility
• Key Features
• Install
• Configuration
• Quick Start
• Permissions & Data Security
• Troubleshooting & Rollback
• Development & Testing
• License

---

🎯 Overview

What problem does it solve?

When users connect custom third-party gateways (e.g. OneAPI, NewAPI, vLLM, Ollama) to DeepSeek Harness, the standard GET /models discovery endpoint only returns raw model IDs (such as gemini-3.7-flash, deepseek-v4-pro, or mimo-v2.5). It omits critical runtime parameters:

• ❌ Missing context limits causes inaccurate token pressure metering or context overflows.
• ❌ Missing multimodal flags (input: ['text', 'image']) disables image uploads in the conversation interface.
• ❌ Missing reasoning levels (reasoningEfforts) hides the thinking intensity slider in chat.

How does this plugin solve it?

1. Auto Discovery & Specification Enrichment: Intercepts llm.discoverModels and automatically fills correct context size and max output tokens.
2. Vision & Thinking Auto-Injection: Intercepts llm.resolveModelInfo and settings.mutate to automatically attach vision input and thinking levels both in runtime and in settings.yaml (~/.dsh/settings.yaml).
3. Zero Official Interference: Only affects custom third-party providers; native DeepSeek channels remain untouched.

---

🧭 Compatibility

Environment	Supported Versions	Status
DeepSeek Harness (DSH)	0.1.0-rc.1 ~ mainline	✅ Verified (Runtime Compatible)
Cordis Framework	^3.0.0	✅ Verified
Node.js	`^20.0.0
OS	Linux, macOS, Windows	✅ Cross-Platform

---

✨ Key Features

• ⚡ Zero-Config Automation: Works right out of the box when adding custom providers in Web Settings.
• 📏 Accurate Capacities: Automatic 1M context for Gemini 3.7 Flash, 1000K for DeepSeek V4 Pro, 128K for GPT-4o, 256K for MiMo 2.5, etc.
• 👁️ Automatic Vision Activation: Unlocks image uploads and visual reasoning in Web chat for multimodal models (Gemini, Claude, GPT-4o, MiMo, Qwen-VL, etc.).
• 🧠 Thinking Intensity Tiers: Injects reasoning effort gears (off, low, medium, high, max) for reasoning models.
• 💾 Dual-Layer Persistence: Automatically writes clean parameters into settings.yaml on save, and patches legacy placeholder models in memory.
• 🛡️ Fault-Tolerant & Offline Fallback: Built-in 30+ core model definitions, local caching, timeout circuit-breaking, and fuzzy matching for dated model snapshots (-20241120).

---

Install

As a standard DSH Profile Bundle, installation is one simple command:

dsh plugin --profile web add github:cinob/dsh-plugin-custom-provider-enhancer

Note: As a Profile Bundle, the plugin automatically mounts and activates. There is no need to manually insert custom-provider-enhancer in profiles/web/cordis.patch.yml.

Uninstallation

dsh plugin --profile web remove dsh-plugin-custom-provider-enhancer

---

⚙️ Configuration

Optional configuration in $DSH_HOME/profiles/web/cordis.patch.yml:

- id: custom-provider-enhancer
  config:
    # Remote LiteLLM metadata catalog URL
    metadataUrl: https://raw.githubusercontent.com/BerriAI/litellm/main/model_prices_and_context_window.json
    # Remote OpenRouter official model metadata URL
    openrouterMetadataUrl: https://openrouter.ai/api/v1/models
    # Request timeout in milliseconds (default: 6000ms)
    timeoutMs: 6000
    # In-memory cache TTL in milliseconds (default: 1 hour)
    cacheTtlMs: 3600000
    # Fallback context window when model is unknown (default: 128000)
    defaultContextWindow: 128000
    # Fallback max output tokens (default: 4096)
    defaultMaxTokens: 4096

---

⚡ Quick Start

1. Start DSH Web GUI (dsh web).
2. Go to Settings ➔ Models ➔ Add Provider.
3. Fill in your Base URL and API Key, then click Fetch available models.
4. Check desired models and click Adopt, then click Save.
5. All specifications (contextWindow, maxTokens, input, reasoningEfforts) are automatically configured and saved!

---

🔒 Permissions & Data Security

• Network Access: Only requests the user-specified custom endpoint GET /models and public model specification database (github.com/BerriAI/litellm).
• Zero Credential Leaks: API Keys are passed through standard authorization headers only during user-initiated discovery; keys are never logged, forwarded, or stored by this plugin.
• Local Sandbox Safe: Strictly operates in-process through Cordis service hooks (ctx.llm, ctx.settings); creates no subprocesses or arbitrary file modifications.

---

🛠️ Troubleshooting & Rollback

• Changes not reflecting:
  Ensure you click Save in Web Settings. If models were added prior to installing the plugin, opening Settings and clicking Save will auto-enrich them.
• Rollback:
  Run dsh plugin --profile web remove dsh-plugin-custom-provider-enhancer.

---

💻 Development & Testing

# Clone repository
git clone https://github.com/cinob/dsh-plugin-custom-provider-enhancer.git
cd dsh-plugin-custom-provider-enhancer

# Install dependencies
pnpm install

# Run automated tests
pnpm test

# Build distribution bundle
pnpm build

---

📄 License

MIT License