📦 @goodandready/dsh-model-sync

Dynamic LLM Model Catalog Synchronization & Automated Balance Monitoring for DeepSeek Harness

🇬🇧 English • 🇷🇺 Русский • 🇨🇳 中文说明

⚡ Overview

dsh-model-sync keeps your DeepSeek Harness model catalog up-to-date with upstream AI providers.

Instead of manually editing YAML configs whenever a provider launches a new model, adjusts context window limits, or changes pricing, dsh-model-sync automatically discovers new releases, updates capability metadata (vision, tools, reasoning, embeddings), tracks account balances, manages model selection visibility, and reconciles models on the fly without server restarts.

graph LR
    subgraph Trigger [Scheduling & Manual Trigger]
        Cron[⏰ Background Polling Scheduler] --> Engine[dsh-model-sync Core Engine]
        WebUI[🖥️ Settings: Sync Now Button] --> Engine
    end

    subgraph Providers [25+ Upstream Providers]
        Engine --> Registry{Adapter Registry}
        Registry -->|Bearer Auth| P1[OpenAI / DeepSeek / OpenRouter / Groq]
        Registry -->|x-api-key Auth| P2[Anthropic Claude / Custom Gateways]
        Registry -->|query-key Auth| P3[Google Gemini]
        Registry -->|Local Probe| P4[Local Ollama / vLLM / SGLang]
        Registry -->|CommandCode API| P5[Command Code / visibleModels]
        Registry -->|Plan Discovery| P6[ClineBot Subscription Plan]
    end

    subgraph Reconcile [Catalog Reconciler & Auditing]
        P1 --> Normalizer[Model Normalizer & Capabilities Tagger]
        P2 --> Normalizer
        P3 --> Normalizer
        P4 --> Normalizer
        P5 --> Normalizer
        P6 --> Normalizer
        Normalizer --> Diff[Sync Diff Tracker: Added / Deprecated]
        Diff --> Catalog[Active DSH Models Catalog & visibleModels]
    end

    style Trigger fill:#1e1e2e,stroke:#89b4fa,stroke-width:2px,color:#cdd6f4
    style Providers fill:#181825,stroke:#cba6f7,stroke-width:2px,color:#cdd6f4
    style Reconcile fill:#11111b,stroke:#a6e3a1,stroke-width:2px,color:#cdd6f4

✨ Key Features

• 🏷️ Smart Collections & Catalog Auto-Tagging: Automatically categorizes models into standardized tags (coding, reasoning, fast, vision, long-context) during discovery, provides fast pill-button filtering in the Web UI, and introduces a one-click "⭐ Curated Best" selection preset for top flagship models.
• ⚡ Latency TTFT Radar & Sparklines: Probes response latency and Time To First Token (TTFT) across models, renders ASCII sparklines (▂▃▂) and 4-tier speed badges (⚡ < 300ms, 🟢 < 1.2s, 🟡 1.5 - 3.0s, 🔴 > 3.0s), and exposes GET /dsh-model-sync/health/latency for intelligent downstream routing.
• 💸 Dynamic Live Pricing Sync & Prompt Caching: Periodic background updates of live model pricing from OpenRouter and LiteLLM snapshots with persistent disk fallback cache, tracking prompt cache read/write costs (cacheReadPerToken, cacheWritePerToken), with GET /dsh-model-sync/pricing and POST /dsh-model-sync/pricing/sync endpoints.
• 💰 Built-in Reference Pricing Database: Canonical token pricing database automatically enriches models for OpenAI, Anthropic, Groq, Mistral, xAI, and DeepSeek, enabling cost-limit filtering and @cheap auto-resolution even when upstream APIs omit pricing metadata.
• 🎯 Dynamic Smart Virtual Aliases & Auto-Fallback: Route requests to virtual aliases like @fast, @cheap, and @best-code, with automatic fallback chains when circuit breakers open or health probes fail.
• 📏 Real Context Limits & Boundary Safety (contextWindow / maxOutputTokens): precise separation of input and generation limits, prevention of 400 max_tokens errors via auto-clamping, and clear In: 128k | Out: 8k badges in Web UI.
• 🧠 Unified Reasoning & Thinking Configuration: automatic format detection (effort, budget_tokens, boolean) and parameter translation with suppression of incompatible options (such as temperature on o1/o3).
• 🔔 External Diff Webhooks & Event Dispatching: Real-time webhook notifications with HMAC-SHA256 signature verification and Cordis model-sync/diff events when models are added, deprecated, or prices change.
• ⚡ High-Performance Policy Filtering: Fast-path evaluation in filterModels bypasses regular expressions and array filtering when no filtering policy is configured, guaranteeing zero-overhead discovery runs.
• 🔄 One-Click In-App Updates: Built-in update checker in Settings with automatic npm cache bypass on demand (?fresh=1 / Cache-Control: no-cache), ensuring instantaneous discovery of new package releases.
• 🔄 Automated Catalog Discovery: Syncs model lists, aliases, context windows, and capability flags (vision, tools, reasoning, embeddings) across 25+ providers.
• 🛡️ Minijinja Chat Template & Role Compatibility: Automatically sets compat.supportsDeveloperRole: false for Groq open-source models (Qwen, Llama, Mistral, Gemma, DeepSeek) and DeepSeek API, preventing 400 Unexpected message role errors with reasoning models while preserving user overrides.
• 🌐 25+ Built-in Provider Adapters: Pre-configured discovery for OpenAI, Anthropic, Google, DeepSeek, xAI, OpenRouter, Groq, Mistral, Cerebras, Fireworks, HuggingFace, Moonshot, NVIDIA, Qwen, Together, Xiaomi MiMo, SiliconFlow, Command Code, ClineBot, and local Ollama.
• 🎛️ Command Code Model Selection: Full discovery of Command Code models with granular checkbox selection in the UI, automatically syncing visibleModels to llm-commandcode settings.
• 🎯 ClineBot Subscription Plan Sync: Dedicated adapter querying GET /users/me/plan to discover only active subscription-included models, eliminating catalog clutter from out-of-plan endpoints.
• 🔌 Generic / Custom Adapter Support: Connect arbitrary OpenAI-compatible /v1/models endpoints with configurable auth (bearer, x-api-key, query-key, none).
• 📊 Balance & Quota Monitoring: Queries upstream billing endpoints (where supported) to report account credit balances and prevent unexpected exhaustion.
• 📜 Sync History & Diff Auditing: Tracks all newly discovered models, deprecated IDs, and modified capabilities with timestamped diff logs in the Web UI.
• 🖥️ Full Web GUI Panel (Settings → Model Synchronization):
  • Live status cards per provider with active model counters;
  • Instant "Sync Now" button for on-demand catalog refreshes;
  • Granular model checkboxes for selection and visibility management;
  • Provider enable/disable toggles and custom endpoint inputs.

🛠️ Supported Discovery Providers

📦 Quick Installation

dsh plugin --profile web add @goodandready/dsh-model-sync

[!IMPORTANT] Restart DSH Web UI after installation (systemctl --user restart dsh-web) to activate background catalog synchronization.

⚙️ Configuration (settings.yaml)

dsh-model-sync:
  enabled: true
  scheduleEnabled: false
  intervalMinutes: 60
  autoApply: false
  requestTimeoutMs: 15000
  retryAttempts: 3
  concurrency: 4
  providers:
    - provider: commandcode
      enabled: true
      autoApply: false
    - provider: clinebot
      enabled: true
      autoApply: false

🔍 Active Capability Probing & Verification

• Micro-Probes: Accurately test actual model capabilities rather than relying on declared flags:
  • JSON Schema: Tests structured output generation (response_format: { type: "json_schema" }).
  • Tools / Functions: Tests function call generation and parameter handling.
  • Vision: Tests multimodal processing with lightweight image payloads.
  • Developer Role: Validates developer role prompt acceptance for advanced reasoning models.
• UI Actions: One-click "🔍 Verify Capabilities" on any provider card, plus global "🔍 Verify All Models" and "⚡ Verify New Only" buttons in catalog toolbar.
• Verified Badges: Highlights confirmed features with green indicators (✔ Tools, ✔ Vision, ✔ JSON, ✔ DevRole) persisted across restarts.

🦙 Ollama & Local Engine Hub

• Direct Web Management: Pull, monitor, and manage local Ollama models directly from DSH Web UI without SSH.
• Streaming Pull Progress: Real-time layer download progress bar with speed and transferred byte metrics.
• GPU VRAM Management: Displays active models in VRAM (/api/ps) and provides instant "⏏ Unload from VRAM" (keep_alive: 0) to free GPU resources.
• Architecture & Quantization: Displays quantization levels (Q4_K_M, Q8_0), family, context windows, and offers direct deletion.

📄 License

MIT © GooDAnDReaDY