setup-openrouter

name	setup-openrouter
description	Configure OpenRouter as the cloud-LLM backend for skills in this plugin. Use when a skill needs cloud LLM access and the user wants pay-as-you-go routing across Claude, GPT, Gemini, DeepSeek, Llama, Qwen without managing multiple provider keys.

Setup OpenRouter

Single API key, access to ~every hosted LLM.

Why OpenRouter here

• Unified billing and model catalog — easy to swap cheap↔premium per task.
• Live pricing visible per model; easy to pick cost-optimal for bulk runs.
• OpenAI-compatible API — drop-in with the OpenAI SDK.

Procedure

1. Key: check env for OPENROUTER_API_KEY. If absent, ask the user to set it (point at 1Password if they use op-vault). Never print the key.

2. Endpoint: https://openrouter.ai/api/v1 — set as OPENAI_BASE_URL with the OpenAI SDK.

3. Recommended model aliases (update when prices shift — check openrouter.ai/models):

   Tier	Model id (OpenRouter)	Good for
   Ultra-cheap	deepseek/deepseek-chat, google/gemini-2.0-flash-001	Bulk classification, labeling
   Cheap	anthropic/claude-haiku-4-5, openai/gpt-4o-mini	Classification with higher quality, summarization
   Premium	anthropic/claude-sonnet-4-6, anthropic/claude-opus-4-7, openai/gpt-4o	Taxonomy design, nuanced labeling, residue review
   Embeddings	openai/text-embedding-3-small (direct OpenAI), or local sentence-transformers	Clustering, similarity

   Prices change — always fetch current pricing before large runs.

4. Test call: send a 1-token completion to verify routing + key.

5. Set safe defaults in a shared config:

   • Max tokens per call.
   • Concurrent request limit (start at 10).
   • Timeout + retry policy (exponential backoff on 429/5xx).

6. Enable prompt caching for providers that support it (Anthropic, OpenAI) — huge savings on repeated system prompts in classification/taxonomy work.

Cost guardrails

• Any skill invoking OpenRouter on >1000 docs must pre-compute an estimate via choose-approach.
• Set a per-run max_spend_usd envelope; abort if projected exceed.
• Log every call's input/output token count and estimated cost to logs/openrouter-calls.jsonl.

Privacy

OpenRouter forwards to the chosen provider. If the corpus is sensitive, prefer setup-local-llm.