multi-model-orchestrator_skill

Overview

This skill orchestrates multi-model collaboration by calling gemini-agent and codex-agent to assist deep code analysis and decision-making. It treats external agents as advisers while you remain the decision-maker and executor. The skill is optimized for cross-file exploration, complex reasoning, and multi-perspective recommendations for a Python Discord bot codebase.

How this skill works

You run a phased workflow: quick local reconnaissance to scope the task, then dispatch gemini-agent for wide scanning and codex-agent for deep reasoning as needed. Haiku (the relay) packages code context and forwards prompts; gemini returns panoramas and facts while codex returns analytic recommendations. Finally you synthesize agent outputs, make decisions, and apply changes yourself.

When to use it

• Deep code analysis: algorithm understanding, architecture mapping, performance bottlenecks
• Large-scope exploration: 5+ files, module dependencies, call chain tracing
• Complex reasoning: scheme evaluation, concurrency safety checks, logical verification
• Multi-perspective decisions: need different analytical angles and a combined judgment
• When you plan to orchestrate parallel agents and then consolidate findings

Best practices

• Always start with a local reconnaissance step (glob/grep) to identify modules and core files
• Choose task type explicitly: facts-only or suggestion-oriented before calling agents
• Limit each agent’s scope to avoid output truncation; split large projects into modules
• Run agents in parallel but read 1–2 core files yourself during waiting periods to avoid duplication
• Treat agent outputs as advisory evidence; validate and decide changes yourself

Example use cases

• Scan the tools/ and providers/ modules with gemini-agent and then ask codex-agent for optimization trade-offs
• Locate O(n^2) hotspots across 10+ Python files and produce a factual table of functions and complexity
• Evaluate concurrency issues: gemini-agent maps call chains, codex-agent assesses race conditions and mitigation options
• Compare refactor proposals: gemini outlines structure, codex proposes prioritized refactor options
• Split a full-project audit into module-level agents to stay under output limits and then synthesize results