# Pi Fusion A multi-model deliberation harness for the [Pi Coding Agent](https://pi.dev). Pi Fusion takes any complex technical question and runs it through a structured deliberation pipeline -- parallel expert panels, comparative analysis, and grounded synthesis -- producing answers that are more thorough and balanced than any single model can achieve alone. Inspired by the [OpenRouter Fusion](https://openrouter.ai/docs/features/fusion) design pattern. Written in Node.js with native ES Modules, requires no build steps, and ships with ready-to-use presets for OpenCode Go subscription models. --- ## How It Works ### 5x Mode (Full Pipeline) ``` [User Query] | +-------------+-------------+ | | | v v v [Technical] [Devil's] [Systems] Tier 1: Parallel Panel [Expert] [Advocate] [Thinker] (3 models, run in parallel) | | | +-------------+-------------+ | v [Deliberation] Tier 2: Judge [Judge] (Structured JSON comparison) | v [Synthesis] Tier 3: Synthesis [Model] (Grounded final answer) | v [Final Answer] ``` Five LLM calls total. Each panel expert has a different system prompt persona. The Judge produces structured JSON (`consensus`, `contradictions`, `partial_coverage`, `unique_insights`, `blind_spots`). The Synthesis model resolves the contradictions and produces the final answer. ### 3x Mode (Lean Pipeline) Three LLM calls instead of five. Two parallel experts (Technical + Devil's Advocate) followed by one synthesizer that absorbs the Judge and Systems Thinker roles. Roughly 40% cheaper than 5x mode while preserving the core deliberation benefit. --- ## Presets | Preset | Mode | Models | Provider | Use Case | |--------|------|--------|----------|----------| | **GLM-5.2 Fusion** | 3x | All `glm-5.2` | OpenCode Go | Default. Fast, cheap, 1M context. | | **Balanced** | 5x | `kimi-k2.7-code`, `deepseek-v4-pro`, `kimi-k2.6` | OpenCode Go | Best coding quality per LiveBench. | | **Quality / Frontier** | 5x | `opus-4.8`, `gpt-5.5`, `gemini-3.1-pro` | OpenAI | Highest quality, highest cost. | | **Custom** | 5x | User-defined | Any | Full control over every model slot. | The Balanced preset assigns models based on [LiveBench](https://livebench.ai) coding and reasoning averages: - **Technical Expert** and **Synthesis**: `kimi-k2.7-code` (top global coding average) - **Devil's Advocate** and **Judge**: `deepseek-v4-pro` (strong reasoning, critical analysis) - **Systems Thinker**: `kimi-k2.6` (high global average, holistic reasoning) --- ## Getting Started ### Install as a Pi Extension ```bash pi install git:github.com/QuarkOS/Pi-Fusion.git ``` Or from npm: ```bash pi install npm:@quarkos/pi-fusion ``` ### Set Your API Key The default presets use OpenCode Go. Set the API key in your environment: **Windows (PowerShell):** ```powershell $env:OC_GO_CC_API_KEY = "sk-opencode-..." ``` **Linux / macOS:** ```bash export OC_GO_CC_API_KEY="sk-opencode-..." ``` If no OpenCode Go key is found, the client falls back to `OPENAI_API_KEY` and standard OpenAI endpoints. Pi Fusion also reads keys from Pi's own `auth.json` if you have connected a provider through Pi. ### Choose a Preset Inside Pi, run `/fusion-config` to select a preset interactively. From the CLI, run: ```bash npx @quarkos/pi-fusion --setup ``` --- ## Usage ### Inside Pi Once installed, two features are available: - **`/fusion `** -- Runs the full deliberation pipeline and streams the synthesized answer into your Pi session. - **`deliberate` tool** -- Available to the Pi agent itself. When solving complex tasks, the agent can call this tool to get a multi-model deliberation on a sub-problem. ### As a Model Provider Pi Fusion registers itself as a model provider called `fusion`. You can select it as your active model in Pi's model picker, and every message you send will go through the deliberation pipeline automatically. The synthesis streams token-by-token, and any generated code is routed through a file-agent that saves files to disk using Pi's `write` tool. ### Command Line Submit a one-off query: ```bash npx @quarkos/pi-fusion "Explain the tradeoffs between microservices and monoliths" ``` Interactive REPL mode: ```bash npx @quarkos/pi-fusion --interactive ``` Verbose mode (shows individual panel responses): ```bash npx @quarkos/pi-fusion "Write a thread-safe singleton in Go" --verbose ``` Override specific model slots: ```bash npx @quarkos/pi-fusion "Test query" --models "judge=glm-5.2,synthesis=deepseek-v4-pro" ``` --- ## Architecture ``` index.js Pi extension entry point (provider, commands, tools) bin/pi-harness.js Standalone CLI with setup wizard lib/api.js OpenAI-compatible streaming API client with retry logic lib/deliberation.js Deliberation orchestrator (3x and 5x modes) lib/ui.js Terminal formatting for CLI output ``` The API client streams all responses to prevent gateway timeouts, accumulates token usage across calls, handles Kimi-specific temperature constraints (`temperature: 1.0`), and retries transient network errors with exponential backoff. The file-agent step after synthesis uses `deepseek-v4-flash` (cheap, near-unlimited on OpenCode Go) with tool-calling to decide what files to save. --- ## Configuration Pi Fusion looks for a `pi-harness.config.json` in your working directory. If none exists, it uses built-in defaults. The config file is generated automatically when you run the setup wizard. Key fields: | Field | Description | |-------|-------------| | `provider` | Which provider to use (`opencode-go`, `openai`) | | `mode` | `3x` (lean) or `5x` (full pipeline) | | `providers..defaultModels` | Model IDs for each pipeline slot | | `panel..systemPrompt` | Custom system prompts per expert | | `judge.systemPrompt` | Judge comparison instructions | | `synthesis.systemPrompt` | Final synthesis instructions | | `fileAgentModel` | Model used for the file-saving step (default: `deepseek-v4-flash`) | --- ## Development ```bash git clone https://github.com/QuarkOS/Pi-Fusion.git cd Pi-Fusion npm install ``` Run the full test suite (requires a valid `OC_GO_CC_API_KEY`): ```bash npm test # Live API deliberation test npm run verify:stream # Offline stream protocol tests ``` --- ## License [MIT](LICENSE)