Gemini Router

Claude Gemini Router

繁體中文 | 简体中文

A self-evolving Claude Code plugin that intelligently routes tasks to Google's Gemini CLI based on task characteristics, saving Claude tokens for large-context operations.

The more you use it, the smarter it gets — the system remembers your preferences and continuously evolves its decision logic.

Features

• Smart task scoring — automatically evaluates whether a task is better suited for Gemini (large file analysis, summarization) or Claude (code editing, iterative workflows)
• Preference learning — remembers your choices via an experience log, auto-routing familiar task types without asking
• Dissatisfaction detection — watches for negative feedback and learns to avoid routing similar tasks to Gemini in the future
• Self-evolution — implements the same Experience-driven Lifelong Learning (ELL) principles as the AutoSkill research

---

Self-Evolving Learning Loop

User Request → Task Scoring → Experience Log Lookup
                                    ↓
                          ┌─ AUTO_APPROVE → Delegate to Gemini directly
                          ├─ AVOID → Claude handles it
                          └─ Unknown → Ask user
                                              ↓
                                    ┌─ Remember & use → Write AUTO_APPROVE
                                    ├─ Just this once → No record
                                    └─ Claude handles → No record
                                              ↓
                          User dissatisfied? → Write AVOID, auto-skip in future

How This Is "Self-Evolving"

This skill implements the same core principles as AutoSkill (Yang et al., 2026, arXiv:2603.01145) — an Experience-driven Lifelong Learning (ELL) framework:

AutoSkill Concept	Gemini Router Implementation
Experience Detection — monitor interactions for stable preferences	Scoring system detects task characteristics (file size, read/write, complexity)
Pattern Extraction — extract reusable skills from feedback	When user chooses "remember", task pattern is written as AUTO_APPROVE entry
Version Evolution — skills update with continued feedback	AVOID can override AUTO_APPROVE, decision logic continuously self-corrects
Skill Application — apply learned skills to future tasks	Tasks matching AUTO_APPROVE are delegated instantly, zero confirmation

Key shared principles:

• Human oversight — both systems ask user consent before recording, never act unilaterally
• Bidirectional learning — learns from both success (AUTO_APPROVE) and failure (AVOID)
• Persistent memory — experience survives across conversations, never "use and forget"
• Progressive evolution — decision accuracy improves with each use

---

Prerequisites

• Claude Code installed and authenticated
• Gemini CLI installed (npm install -g @google/gemini-cli)

Installation

Step 1: Add the marketplace

In Claude Code, run:

/plugin marketplace add audi0417/claude-gemini-router

Step 2: Install the plugin

/plugin install gemini-router@claude-gemini-router

Or use the interactive plugin manager:

/plugin

Then navigate to Discover tab and select gemini-router.

Alternative: Manual install

Copy the skill files directly:

git clone https://github.com/audi0417/claude-gemini-router.git
cp -r claude-gemini-router/plugins/gemini-router/skills/gemini ~/.claude/skills/gemini

How It Works

1. Evaluate — When you give Claude a task, the skill scores it for Gemini suitability
2. Decide — High-scoring tasks prompt you to choose: delegate to Gemini, or keep with Claude
3. Learn — Your choices are logged so similar tasks are auto-routed next time
4. Adapt — If Gemini's output is unsatisfactory, the skill learns to avoid that task type

Scoring Criteria

Signal	Weight
Large files (>50KB total)	+2
Read-only / analysis task	+2
Matches AUTO_APPROVE pattern	+2
User explicitly requests Gemini	+2
Multi-file cross-reference	+1
Summarization / extraction	+1
Token conservation benefits	+1

Disqualifiers (always use Claude)

• Matches an AVOID pattern in experience log
• File writing or editing
• Iterative tool-use loops
• Short, simple tasks

Experience Log (experience.md)

All learning outcomes are stored in experience.md — transparent and human-readable:

### [AUTO_APPROVE] 2026-04-08
**Pattern**: Full Go codebase read-only analysis
**Trigger keywords**: codebase, analysis, review
**Reason**: Large file volume, read-only, Gemini 1M context is a perfect fit

### [AVOID] 2026-04-08
**Pattern**: File writing, editing, or code generation
**What went wrong**: Gemini headless is single-shot, cannot iterate on edits
**Lesson**: Any task requiring Write/Edit tool use stays with Claude

---

Usage Examples

Auto-delegation (learned pattern)

You: Analyze the architecture across these 20 files
Claude: Routing to Gemini (auto-approved: large codebase analysis)
[Gemini Response] ...

First encounter (asks your preference)

You: Read this 200KB log file and find the errors
Claude: This task looks like a good fit for Gemini (large context). How should I handle it?
  1. Use Gemini & remember this task type
  2. Use Gemini, just this once
  3. Claude handles it

Dissatisfaction feedback

You: Gemini's analysis was off, redo this yourself
Claude: Should I mark this task type as a poor fit for Gemini?
  1. Yes, log it
  2. No, just this once

---

License

MIT