Problem

Currently, all custom commands defined in or must go through the AI model:

This consumes input/output tokens even for simple, deterministic commands like:

• • just runs a local script and returns output
• • displays system information
• Custom info queries that don't need AI processing

Example Cost

A simple command consumes ~100-200 tokens just to:

1. Parse the command
2. Execute tool to run script
3. Format the response

This adds up for frequently-used commands.

Proposed Solution

Add a flag (or similar mechanism) for custom commands:

When :

1. Command is intercepted at gateway level (like native )
2. Tool is executed directly
3. Output is sent to user without entering AI session
4. Zero tokens consumed for the command itself

Benefits

• 💰 Significant cost savings for frequent info commands
• ⚡ Faster response times (no model latency)
• 🏗️ Clearer architecture: AI commands vs system commands
• 🔄 Backward compatible (default )

Alternatives Considered

• Keep current behavior (always through AI) - works but wasteful
• Use hooks - not suitable for user-triggered commands
• Separate command system - adds complexity

Related

• Feature: Plugin callback handler for Telegram inline buttons (bypass LLM) #4280 - Plugin callback handler (bypass LLM for callbacks)
• Feature: Gateway-level /switch command to change auth accounts without AI #7533 - Gateway-level command for auth switching
• Feature Request: Tool execution hook events (tool:before, tool:after) #7597 - Tool execution hook events