OpenCode Orchestrator

Autonomous Multi-Agent Orchestration Engine for Software Engineering

⚡ Quick Start

npm install -g opencode-orchestrator

In an OpenCode environment:

/task "Implement"

Overview

OpenCode Orchestrator is a Distributed Cognitive Architecture designed for high-precision software engineering. It operates on a strict "Verify, then Trust" philosophy, distinguishing itself from simple stochastic chatbots by enforcing rigorous architectural standards.

The system is a testament to the operational paradox: Complexity is easy; Simplicity is hard.

While the user interaction remains elegantly minimal, the internal architecture encapsulates a rigorous alignment of microscopic state management (Rust atoms) and macroscopic strategic planning (Agent Topology). Every component reflects a deep design philosophy aimed at abstracting chaos into order.

Building this system reaffirmed a timeless engineering truth: "Simple is Best" is the ultimate complexity to conquer. This engine is our answer to that challenge—hiding the intricate dynamics of Autonomous Agentic Collaboration behind a seamless, user-friendly veil.

This philosophy extends to efficiency. We achieved Zero-Configuration usability while rigorously optimizing for performance—delivering higher quality outcomes than alternatives while saving ~40% of tokens. By maximizing the potential of cost-effective models like GLM-4.7, we prove that superior engineering—not just raw model size—is the key to autonomous performance.

📊 Workflow

              [ User Task Input ]
                        │
            ┌───────────▼───────────┐ <────────────────────────────────────────┐
            │   🫡 COMMANDER (Hub)  │  (Orchestration)                         │
            └───────────┬───────────┘                                          │
                        │                                                      │
            ┌───────────▼───────────┐                                          │
            │   🗓️ PLANNER (Map)    │  (Create TODO.md)                        │
            └───────────┬───────────┘                                          │
                        │                                                      │
     ┌──────────────────▼──────────────────┐                                   │
     │   ⚡ COMMANDER: Parallel Spawning   │                                   │
     └──────┬───────────┬───────────┬──────┘                                   │
            │           │           │                                          │
     ┌──────▼───┐ ┌─────▼────┐ ┌────▼─────┐                                    │
     │ 🔨 WORKER│ │ 🔨 WORKER│ │ 🔨 WORKER│                                    │
     └──────┬───┘ └─────┬────┘ └────┬─────┘                                    │
            │           │           │                                          │
     ╔══════▼═══════════▼═══════════▼══════╗                                   │
     ║   🔍 COMMANDER: Parallel Reviewers  ║                                   │
     ╚══════╤═══════════╤═══════════╤══════╝                                   │
            │           │           │                                          │
     ┌──────▼───┐ ┌─────▼────┐ ┌────▼─────┐                                    │
     │🔍REVIEWER│ │🔍REVIEWER│ │🔍REVIEWER│                                    │
     └──────┬───┘ └─────┬────┘ └────┬─────┘                                    │
            │           │           │                                          │
           ═▼═══════════▼═══════════▼═                                         │
           │      🚦 SYNC BARRIER      │                                         │
           ═════════════╤═════════════                                         │
                        │                                                      │
            ┌───────────▼───────────┐                                          │
            │ ✔️ MASTER REVIEWER    │  (E2E Verification)                      │
            └───────────┬───────────┘                                          │
                        │                                                      │
              __________▼_________                                             │
             ╱                    ╲    NO (Loop / Auto-Correction)             │
            ╱   ✅ All TODOs?      ╲ ──────────────────────────────────────────┘
            ╲   🛡️ Error Rate 0%?  ╱
             ╲____________________╱
                        │ YES
                        │
                [ 🎖️ MISSION SEALED ]

🧠 Cognitive Architecture & Key Strengths

📉 Adaptive Context Gating (EMA-based)

We combat "Context Drift" using a mechanism derived from Exponential Moving Average (EMA) algorithms. Irrelevant conversation noise follows a rapid decay curve, while critical architectural decisions are reinforced into Stable Core Memory. This functions as an Attention Sink, allowing agents to work indefinitely without Catastrophic Forgetting.

🧬 BDI (Belief-Desire-Intention) Collaboration

The system implements a variant of the BDI Software Agent Model:

• Belief (Context): Shared state & file system reality.
• Desire (Mission): The user's goal (e.g., "Fix this bug").
• Intention (Plan): The TODO.md roadmap execution. Agents do not merely "chat"; they collaborate to align their Beliefs with Desires through strictly executed Intentions, mirroring human engineering squads.

⚙️ Neuro-Symbolic Hybrid Engine

Pure LLM approaches are stochastic. We bind them with a Neuro-Symbolic Architecture that anchors probabilistic reasoning to the deterministic precision of Rust-based AST/LSP Tools. This ensures every generated token is grounded in rigorous syntax analysis, delivering high performance with minimal resource overhead.

⚡ Dynamic Fork-Join Parallelism with Backpressure

The engine features an Intelligent Load-Balancing System that fluidly switches between synchronous barriers and asynchronous Fork-Join patterns. It monitors System Backpressure to dynamically adjust concurrency slots in real-time (Adaptive Throttling), maximizing throughput on high-end hardware while maintaining stability on constrained environments.

🎯 Iterative Rejection Sampling (Zero-Shot Defense)

We employ a Rejection Sampling Loop driven by the Reviewer Agent (Reward Model). Through the Metric-based Strict Verification Protocol (MSVP), code paths that fail execution tests are pruned. The system iterates until the solution converges on a mathematically correct state (0% Error Rate), rejecting any solution that lacks evidence.

🧩 Externalized Chain-of-Thought (CoT)

The Planner's TODO.md serves as an Externalized Working Memory (Scratchpad). This persistent Symbolic Chain-of-Thought decouples detailed planning from the LLM's immediate context window, enabling the orchestration of massive, multi-step engineering tasks without logical degradation.

⚡ Agents

Developer's Note

Full Developer's Note →

System Architecture →

📄 License

MIT License. See LICENSE for details.