External Model Delegation Pattern

Multi-model orchestration routes prompts to different models automatically by using a UserPromptSubmit hook to classify each prompt into quality and cost tiers, then injecting a delegation command into additionalContext for Claude to execute. This ensures optimal model selection without manual switching.

Optimizing LLM costs for code reviews and refactoring is best handled by tiered delegation, which classifies prompts across six model tiers like QWEN, DEEPSEEK_FLASH, and CLAUDE. This cost optimization approach matches task complexity to the cheapest capable model, preventing cost blowups on simple tasks.

Set up Claude Code hooks for multi-model orchestration by configuring a UserPromptSubmit hook to intercept prompts, classify them into model tiers, and inject delegation instructions into additionalContext. You must also install delegation scripts in ~/bin/ that accept prompt arguments, support required model flags, and write responses to stdout.

You can use prompt routing with DeepSeek and Kimi models for CLI scripting by configuring delegation scripts that support specific flags like deepseek flash, deepseek pro, and kimi quiet. The scripts must accept the prompt as the first argument, output results to stdout, and return correct exit codes.

Multi-model delegation requires delegator scripts to write results to stdout and use exit codes to signal success or failure. This deterministic delegation script contract ensures the hook-injected context can reliably capture the external model's response and return it to the user.

Prompt routing fails to execute the delegated model script when delegation scripts are not installed in the expected ~/bin/ directory, lack the required model flags, or do not write responses to stdout with correct exit codes. The UserPromptSubmit hook relies on this strict contract to route prompts successfully.