@quarkos/pi-fusion

Multi-model deliberation harness (OpenRouter Fusion pattern) for OpenCode Go

Packages

Package details

package

Install @quarkos/pi-fusion from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@quarkos/pi-fusion
Package
@quarkos/pi-fusion
Version
1.0.1
Published
Jun 15, 2026
Downloads
not available
Author
quarkos
License
MIT
Types
package
Size
45.9 KB
Dependencies
3 dependencies · 0 peers

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

Pi Fusion

Pi Fusion is an orchestration harness and agent extension that implements a multi-model deliberation pipeline. Inspired by the OpenRouter Fusion design pattern, it takes any complex technical request or coding query and runs it through a three-tier process: parallel expert panels, a structured comparative analysis, and a final grounded synthesis.

It is written in Node.js with native ES Modules, requires no build steps, and features out-of-the-box configuration for OpenCode Go subscription models.


How It Works

The deliberation pipeline consists of three sequential stages:

                  [User Query]
                       |
         +-------------+-------------+
         |             |             |
         v             v             v
    [Technical]    [Devil's]     [Systems]       Tier 1: Parallel Panel
     [Expert]     [Advocate]    [Thinker]        (System Prompt Personas)
         |             |             |
         +-------------+-------------+
                       |
                       v
               [Deliberation]                    Tier 2: Judge
                  [Judge]                        (Structured JSON output)
                       |
                       v
                 [Synthesis]                     Tier 3: Synthesis
                   [Model]                       (Grounded final answer)
                       |
                       v
                [Final Answer]

Tier 1: Panel (Parallel Execution)

The user query is sent to three separate expert models in parallel:

  • Technical Expert (qwen3.7-plus): Evaluates correctness, architectural patterns, performance, and security.
  • Devil's Advocate (deepseek-v4-pro): Challenges assumptions, identifies edge cases, highlights risks, and evaluates simpler alternatives.
  • Systems Thinker (glm-5.1): Focuses on integration, API design, testing strategies, long-term technical debt, and maintainability.

Tier 2: Judge (Deliberative Analysis)

A comparison model (qwen3.7-plus) reviews the panel responses to find agreements and conflicts. It produces a structured JSON output with five keys:

  • consensus: Core technical decisions where the experts agree.
  • contradictions: Specific design conflicts or tradeoffs.
  • partial_coverage: Points raised by some but not all models.
  • unique_insights: Non-obvious optimizations or approaches.
  • blind_spots: Critical omissions or risks that none of the models addressed.

Tier 3: Synthesis (Final Grounded Answer)

A final model (qwen3.7-plus) synthesizes the user query, the panel responses, and the Judge's structured JSON analysis into a comprehensive markdown answer.


Configuration and Setup

  1. Install dependencies in your project directory:

    npm install
    
  2. Set your OpenCode Go API key:

    • Windows (PowerShell):
      $env:OC_GO_CC_API_KEY="sk-opencode-..."
      
    • Linux/macOS:
      export OC_GO_CC_API_KEY="sk-opencode-..."
      

If no OpenCode Go key is found, the client looks for a standard OPENAI_API_KEY and falls back to standard OpenAI endpoints.


Command Line Usage

You can run Pi Fusion directly as a command-line tool.

Direct Run

Submit a query from the terminal:

node bin/pi-harness.js "Explain the tradeoffs between microservices and monoliths"

Verbose Mode

Use the --verbose or -v flag to inspect the individual panel responses:

node bin/pi-harness.js "Write a thread-safe singleton pattern in Go" --verbose

Interactive Mode (REPL)

Launch a persistent chat session to run multiple queries:

node bin/pi-harness.js --interactive

Model Overrides

You can override default models for any tier using the --models flag:

node bin/pi-harness.js "Test query" --models "technical_expert=deepseek-v4-pro,judge=glm-5.1"

Installing as a Pi Agent Extension

Pi Fusion is designed to be fully compatible with the Pi Coding Agent (pi.dev). When installed, it adds a /fusion command and a deliberate tool to the agent.

Install the extension from GitHub:

pi install git:github.com/QuarkOS/Pi-Fusion.git

Or install it locally from your project folder:

pi install .

Registered Features inside Pi:

  • Slash Command: /fusion <prompt> — Runs the multi-model deliberation pipeline directly in your Pi terminal session.
  • Agent Tool: deliberate — Allows the Pi Coding Agent to call this deliberation process programmatically when solving complex coding problems.