Skip to content

budagov-lab/DreamTeam

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

56 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DreamTeam — Autonomous Development Cruiser for Cursor

License: PolyForm System: Cursor

A long-range Autonomous Development Cruiser for Cursor capable of executing 500+ sequential tasks without quality degradation. Built for fault tolerance, continuous learning, and multi-layered agent orchestration using ping-pong execution loop.

Important

Dispatcher Architecture: DreamTeam is designed to offload deep work from the main chat. The Dispatcher coordinates Left and Right Orchestrators to run batches of 15+ tasks with minimal supervision. Each context switch (ping-pong) performs a context reset, ensuring the dispatcher never hits context ceilings. This is the key to executing 500+ tasks with zero performance degradation.

Quick Start:

  1. python -m dreamteam new-project . (in an empty folder)
  2. Open in Cursor → /start + your goal
  3. Start or resume the execution loop: /run

Pipeline: High-Performance Autonomy

The system uses a recursive dispatching loop. The Dispatcher coordinates specialized orchestrators to handle batches of tasks, keeping the main context clean and stable.

---
config:
  layout: fixed
  look: handDrawn
---
flowchart BT
 subgraph PhaseP["Planning"]
    direction TB
        P2["Sub-Planner"]
        P1["Planner"]
  end
 subgraph PhaseE["Execution Loop"]
    direction TB
        E2["Reviewer"]
        E1["Developer"]
        E3["DevExperience"]
        E4["Git-Ops"]
        Upd["update-task Done"]
  end
 subgraph PhaseM["Maintenance"]
    direction TB
        M2["FixPlanner"]
        M1["Learning / Meta"]
        M3["Researcher"]
        M4["Auditor"]
  end
 subgraph Ops["Operational Phases"]
    direction LR
        PhaseP
        PhaseE
        PhaseM
  end
 subgraph Context["Isolated Agent Context"]
    direction TB
        Ops
        LR_Agent{"Left and Right Orchestrator"}
        Term[["Terminal Subagent"]]
  end
 subgraph Engine["DreamTeam Cruiser Engine"]
    direction LR
        DP["Dispatcher"]
        Context
        DAG[("Task DAG")]
        RAG[("Memory DB")]
        Counter[("Counter")]
  end
    User(["User Goal"]) --> DP
    DP -- Switch batch --> LR_Agent
    LR_Agent --> Ops
    P1 --> P2
    E1 --> E2
    E2 --> E3
    E3 --> E4
    E4 --> Upd
    M1 --> M2 & E1
    Ops -.-> Term
    P2 --- DAG
    M2 --- DAG
    M3 --- RAG
    Upd -. "Batch Limit =15 or
    Context Overflow" .-> DP
    P2 -. Planning Done .-> DP
    M4 --> RAG
    Counter --> M4 & M3 & M1
    DAG --> M4 & M3

     P2:::clsWorker
     P1:::clsWorker
     E2:::clsWorker
     E1:::clsWorker
     E3:::clsWorker
     E4:::clsWorker
     Upd:::clsWorker
     M2:::clsMaintain
     M1:::clsMaintain
     M3:::maintain
     LR_Agent:::clsOrch
     Term:::clsInfra
     DP:::clsMain
     DAG:::clsDB
     RAG:::clsDB
    classDef clsMain fill:#4f46e5,color:#fff,stroke:#3730a3,stroke-width:2px,rx:10
    classDef clsOrch fill:#f8fafc,stroke:#94a3b8,stroke-width:2px,stroke-dasharray: 4,rx:10
    classDef clsWorker fill:#ffffff,stroke:#e2e8f0,stroke-width:1px,rx:4
    classDef clsMaintain fill:#ecfdf5,stroke:#10b981,color:#064e3b,rx:4
    classDef clsInfra fill:#0f172a,color:#fff,stroke:#1e293b,rx:2
    classDef clsDB fill:#fff,stroke:#64748b,stroke-width:2px,shape:cylinder
    style LR_Agent stroke:#00C853,fill:#FFD600
    style Context fill:#f8fafc,stroke:#e2e8f0,stroke-width:1px
    style DAG stroke:#000000,fill:#FF6D00
    style RAG fill:#FF6D00
    style Counter fill:#FF6D00
    style User fill:#00C853
    style Engine fill:#fdfaff,stroke:#c4b5fd,stroke-width:2px
Loading

Economy & Model Selection

Important

Model Inheritance: By default, all sub-agents inherit the model name chosen in the main chat (where /start or /run was invoked).

To optimize project economy:

  • Heavy Reasoning (Planner, Auditor, Researcher, Learning): We recommend using frontier models (e.g., Claude 4.6 Sonnet) for high architectural compliance.
  • Routine Tasks (Developer, Reviewer, Git-Ops): These can often be shifted to more economical models to maintain sustainability over 500+ sequential tasks.

Thoughtful model selection ensures the Project Budget lasts for the entire Cruiser journey.


Under the Hood: Scalable Autonomy

The system is built to minimize "Main Chat" context overflow. Using a Dual Orchestrator system (Left/Right), DreamTeam offloads execution to sub-agents, leaving the main chat lean and responsive. This architectural split allows massive task sequences to run even on non-frontier models.

AI Sub-Agent Hierarchy

DreamTeam uses a multi-layered intelligence system to ensure stability over long durations:

  1. Level 1: Fleet Control (Dispatcher): The entry point. It doesn't perform tasks but manages the switching between "Left" and "Right" Orchestrators. This ensures that even for 1000-task journeys, the main chat context remains lean and responsive.
  2. Level 2: Task Orchestration (Orchestrators): Specialized agents that run inside a fresh context. They decide whether to launch the Planning Phase or the Execution Loop and handle all self-correction triggers.
  3. Level 3: Specialized Workers:
    • Planner & Sub-Planner: Decompose high-level goals into a detailed task DAG.
    • Developer: Implements features and runs tests.
    • Reviewer: Verifies code quality and architectural compliance.
    • Git-Ops: Handles commits and repository maintenance.
    • Maintenance Agents: (Researcher, Learning, Meta-Planner, Auditor) Keep the context compressed and the pipeline optimized.

Core Mechanisms

Fault Tolerance — Nothing Gets Lost

The system is designed to recover from crashes, mismatches, and stuck tasks without manual intervention:

  • run-next: Verifies DB↔Files consistency, auto-syncs if needed, and resets stuck tasks.
  • recover: Full system reset, integrity verification, and memory health check.
  • State-in-DB: All state lives in SQLite. The Cruiser can resume after a break without losing a single bit of context.

Learnability — The Pipeline Adapts

DreamTeam improves from production feedback instead of degrading:

  • DevExperiencer: Records every outcome, attempt count, and time spent.
  • Learning Agent: Analyzes the Experience DB to detect patterns of failure or high friction.
  • FixPlanner: Automatically adjusts upcoming tasks (library choices, dependency updates) to avoid recurring roadblocks.
  • Developer Updates: The system may augment .cursor/agents/developer-addendum.md with additional instructions to permanently adopt successful patterns.

Analytics Dashboard — Monitor the Friction

Launch a minimalistic web dashboard to track your Cruiser's performance:

  • KPIs: Total tasks, estimated tokens, and Friction Score (Avg Attempts).
  • Visualization: Identify hallucination spikes and time-heavy tasks.
  • Task Lineage: Track original plans vs. tasks added during self-correction.

Command: python -m dreamteam dashboard


Documentation


License

MIT license


Crafted for Cursor adepts with love from BuLab

About

Autonomous system that runs hundreds of development tasks sequentially without losing quality. Designed for developers and teams to speed up projects and integrate seamlessly with Cursor.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors