Skip to content

phase-router

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

phase-router

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
config.json
config.json
 
 
extension.mjs
extension.mjs
 
 

README.md

🔄 Phase Router

LLM-driven intent detection + intelligent model switching — a Copilot CLI Extension

The AI agent semantically classifies the user's task intent (research, coding, review…) and switches to the optimal model via the set_phase tool. Inspired by oh-my-openagent's Sisyphus multi-agent intent classification system.

Core idea: Different tasks deserve different models. Use a fast, cheap model for exploration (GPT-5.4 mini), a code-specialized model for implementation (GPT-5.3 Codex), and a powerful reasoning model for architecture design (Opus). Phase Router injects phase context via additionalContext, letting the agent decide when to switch — works with prompts in any language.

Cross-family strategy: The default config intentionally mixes Claude and GPT models. Using GPT to review Claude-generated code (and vice versa) catches blind spots that same-family review would miss.

✨ Features

  • ✅ LLM-driven classification — the agent decides intent semantically, no keyword matching, works in any language
  • ✅ Intelligent model switching — automatically selects the best model for each task type
  • ✅ Cross-family review — GPT reviews Claude code, different training data = different perspectives
  • ✅ Manual override — set_phase tool for explicit phase control
  • ✅ Fully configurable — customize model per phase
  • ✅ Persistent config — settings saved to ~/.copilot/phase-router.json

📊 Task Phases

Phase Emoji Default Model Description
research 🔍 gpt-5.4-mini Investigation, exploration, learning
planning 📋 claude-opus-4.6 Architecture design, system planning
coding 💻 gpt-5.3-codex Implementation, feature building
refactoring ♻️ claude-sonnet-4.6 Code restructuring, cleanup, optimization
review 🔎 gpt-5.4 Code review, security audit, quality check
debugging 🐛 gpt-5.3-codex Bug fixing, error diagnosis, troubleshooting
general 💬 claude-sonnet-4.6 General conversation, other tasks

📦 Installation

Project-level (per project)

cp -r .github/extensions/phase-router your-project/.github/extensions/

User-level (global)

cp -r .github/extensions/phase-router ~/.copilot/extensions/

Restart Copilot CLI or run /clear to load.

🚀 Usage

Just type your prompt normally. The agent will semantically determine whether to switch phases:

> explain how the auth middleware works
  → Agent detects Research intent, calls set_phase("research") → gpt-5.4-mini

> implement user registration API
  → Agent detects Coding intent, calls set_phase("coding") → gpt-5.3-codex

> review this PR for security vulnerabilities
  → Agent detects Review intent, calls set_phase("review") → gpt-5.4

You can also explicitly ask the agent to switch:

switch to planning phase
show current phase config
set the review phase to use claude-opus-4.6

⚙️ Configuration

Field Description
enabled Enable/disable phase switching
showNotifications Show/hide phase switch notifications
defaultPhase Default phase on session start
phases.{name}.model Model to use for that phase

User Override

Create ~/.copilot/phase-router.json with only the fields you want to change:

{
  "phases": {
    "coding": { "model": "claude-sonnet-4.6" },
    "review": { "model": "claude-opus-4.6" }
  }
}

🔧 How It Works

  1. User submits a prompt
  2. onUserPromptSubmitted hook fires
  3. additionalContext is injected with current phase info and available phases
  4. Agent semantically evaluates whether intent has changed — if so, calls set_phase
  5. session.setModel() switches the model (takes effect on the next turn)

🛠️ Available Tools

Tool Description
set_phase Switch task phase — triggered by agent or user
show_phase Show current phase status and model configuration
configure_phase Update a phase's model and persist to config file

💡 Model Selection Rationale

Task Type Recommended Tier Why
Exploration / research Fast & cheap (GPT-5.4 mini, Haiku) Speed over quality, low cost
Implementation / debugging Code-specialized (GPT-5.3 Codex) Optimized for tool-driven coding workflows
Refactoring / general Balanced (Sonnet) Good quality-to-cost ratio
Architecture / review Premium (Opus, GPT-5.4) Requires strongest reasoning capabilities

⚠️ Known Limitations

  • session.setModel() takes effect on the next turn, not the current one
  • Phase switching relies on the agent's semantic understanding (all modern LLMs handle this well)
  • State resets to default phase on session restart or /clear