[Skill proposal] `hybrid-agent`

[sdevinenamd]

Proposed skill name

hybrid-agent

Does something like this already exist?

Not sure — happy to discuss

Where should this skill live?

Path A: incubated in this repo (may move to a product repo and continbue to be registered here)

Catalog focus area

Application integration

Skill description

Hybrid Local-Cloud Agent Skill

User Goal

Enable cost-effective, privacy-first AI work that combines local inference with optional cloud offloading (Fireworks) while maintaining memory continuity across both environments.

Platform Scope

Client: WSL/ Ubuntu 24.04 (agent runtime)

Server: Windows host (Lemonade inference server)

Cloud: Fireworks API (optional offload for heavy workloads)

Inputs and Outputs - Example

Inputs

Hand-drawn UI images (PNG/JPG): Sketches, wireframes, mockups
User preferences via memory: layout style, color schemes, framework choice
Model selection command: /model local or /model cloud-kimi
Configuration: lemonade endpoint (http://172.18.0.1:13305/api/v1), Fireworks API key

Outputs

Working website code: HTML, CSS, JavaScript files
Memory entries:
MEMORY.md: "User prefers card-based layouts", "Uses Tailwind CSS"
USER.md: Design preferences, accessibility requirements
Skills created:
ui-from-sketch.md: Vision → website workflow
offload-to-cloud.md: Model switching procedure
Session continuity: Memory persists across local ↔ cloud switches

Process Flow

1. User uploads sketch → Hermes receives image
2. vision_analyze (local qwen multimodal) → extract UI structure
3. Generate HTML/CSS/JS → save to workspace
4. Update memory → "User likes minimalist design"
5. Heavy task detected → /model cloud-kimi → continue in cloud
6. Task complete → /model local → return to local
7. Final output + enriched memory for future sessions