Offload Tasks to LM Studio Models

**Judgement**: Strong, well-thought-out skill for environments that already run LM Studio 0.4+ locally. Very useful for cost-cutting and privacy-preserving workflows, but entirely dependent on a correctly configured local LM Studio server and Node tooling. **What it does** - Probes a local LM Studio server (`http://127.0.0.1:1234`) and lists available models. - Selects an appropriate model based on task (vision, embeddings vs text), context length, and whether instances are already loaded. - Uses a Node script (`lmstudio-api.mjs`) to send chat requests, supporting temperature, max tokens, and multi-turn state via `response_id`. - Optionally loads/unloads models via the LM Studio REST API and verifies unload by checking `loaded_instances`. **Key strengths / advantages** - Built explicitly to offload “cheap” or repetitive work (summaries, extraction, classification, rewriting, brainstorming) from paid APIs to local models. - Provides a clear, stepwise workflow: preflight check → list models → choose model → (optional) load → chat → (optional) unload. - Handles edge cases like multiple instances per model key, and warns not to confuse `model key` with `instance_id`. - Includes retry logic for transient failures and guidance for handling “model not found”, server errors, and memory issues. - Supports stateful conversations using `response_id` / `previous_response_id`, not just one-shot calls. **Risks / limitations** - Hard dependency on a running LM Studio server at `127.0.0.1:1234` with `Authorization: Bearer lmstudio`; if that service isn’t up or is misconfigured, the skill fails. - Requires Node.js (or compatible environment) for the provided scripts; in constrained or sandboxed environments, `exec` calls may be limited. - Quality and speed depend entirely on the user’s chosen local models and hardware; for complex reasoning, results may be noticeably worse than strong paid models. - More moving parts than a direct API call (preflight, model selection, load/unload), which can introduce operational complexity. **Recommended scenarios** - Users already running LM Studio locally, who want to systematically reduce cloud token spend. - Workflows where many tasks are repetitive, low-risk, or tolerant of slightly lower quality (e.g., bulk summarization or tagging). - Privacy-sensitive domains (internal docs, customer data, proprietary code) where keeping data on-device is essential. - Hybrid setups where this skill does first-pass processing locally, and a stronger paid model only handles escalations or final polishing.