v1.8.0 — Competitive Multi-Model Redesign

What Changed

This fork redesigns the entire model routing architecture based on actual benchmark evidence (March 2026).

Before (Original v1.7.74)

Frontend tasks → Gemini (hardcoded)
Backend tasks → Codex (hardcoded)
Claude → Orchestrator only

Problem: No published benchmarks support this static routing. SWE-bench scores are nearly tied (Claude 80.8%, Gemini 80.6%, Codex 80.0%).

After (v1.8.0 Competitive)

Every task → ALL models compete in parallel → Weighted evaluation → Best output wins

Key Changes (25 files, +1048/-592 lines)

🏗️ Architecture

• Competitive dispatch: Every significant task dispatched to Codex + Gemini + Claude (self) in parallel
• Weighted evaluation: Benchmark-informed criteria score each output per task type
• 3 dispatch modes: Competitive (all 3, default), Focused (best-match only), Quick (Claude only)
• Consensus scoring: Review findings tagged with confidence (3/3, 2/3, 1/3 model agreement)

📊 Benchmark Evidence

📝 Files Changed

🚫 What Was Removed

• All "前端模型=Gemini, 后端模型=Codex" static routing references
• All "后端权威/前端权威" (backend authority/frontend authority) trust rules
• Domain-based routing tables in execute.md

✅ What Was Added

• Benchmark evidence table with sources
• Weighted evaluation criteria per task type (analysis, planning, implementation, review, debug)
• Model-specific strengths AND limitations in every prompt
• Consensus scoring with agreement levels in reviews
• Dispatch mode selection (competitive/focused/quick)

Sources

• SWE-bench Leaderboard
• SmartScope LLM Coding Benchmark 2026
• Milvus AI Code Review
• Index.dev Gemini vs Claude Frontend Test
• Render AI Coding Agents Benchmark