SEUFSG/llm-chat

LLM Chat

Route messages to GitHub Copilot, Kimi, MiniMax, or GLM via the llm-gateway MCP server.

Two Modes

Manual Mode — User specifies the model

Use when the user names a specific model or provider, or when precision matters.

1. If provider status is unknown, call llm_status to check authentication
2. If needed, call llm_models to see available models
3. Call llm_chat with mode: "manual", provider, model, and messages
4. Present the response with clear attribution: [Provider / Model Name] followed by the response

Auto Mode — Claude selects the best available model

Use during multi-agent tasks, parallel work, or when the user says "use whatever model is best".

1. Call llm_models to perceive all currently available models (REQUIRED before routing)
2. Identify the task type from context: code_generation, code_review, reasoning, chinese_writing, translation, quick_qa, math, creative, long_context
3. Call llm_chat with mode: "auto", task, and messages
4. The router walks the fallback chain from settings.json and calls the first available model
5. Present the response with attribution showing which model was actually used

Attribution Format

Always show which model answered:

[GitHub Copilot / GPT-4o] [response content here]

Error Recovery

• If llm_chat fails with "not authenticated" → tell user to run llm_login [provider]
• If auto mode fails with "All models in fallback chain exhausted" → show the tried list, suggest the user authenticate a provider or add a routing rule
• If manual mode fails with "model not found" → call llm_models and show the user what's available