DreamTeam Lite (Cursor plugin)

DreamTeam Lite is a Cursor plugin for developers and vibe-coders that helps execution move autonomously. It keeps your goal and task state in sync, routes work between specialized agents, and helps you continue from where you left off after interruptions.

Workflow (Dispatcher is the central switch)

The main chat acts as the Dispatcher: it initializes state.json, then uses last_batch_result and last_finished_orchestrator written by orchestrators to decide which side (Left/Right) runs next. Each orchestrator completes a unit of work and writes fresh batch markers back to state.json, so the loop keeps going without manual babysitting.

Add Plugin Manually (Local)

If you want to add DreamTeam Lite to Cursor without publishing/installing from the marketplace, copy the plugin bundle into Cursor’s local plugins folder.

1. In this repository, locate dreamteam/.cursor-plugin/.
2. Create the folder:
   • %USERPROFILE%\.cursor\plugins\local\dreamteam-lite\
3. Copy the whole dreamteam/.cursor-plugin/ folder into:
   • %USERPROFILE%\.cursor\plugins\local\dreamteam-lite\
4. Verify that plugin.json exists at:
   • %USERPROFILE%\.cursor\plugins\local\dreamteam-lite\.cursor-plugin\plugin.json
5. Restart Cursor (or run Developer: Reload Window).
6. In any workspace, open Cursor chat and test:
   • /start-dreamteam <FULL GOAL TEXT>
   • then /run-dreamteam (if execution was not completed in the first step).

After start/run, runtime state files are created in the current workspace under .dreamteam-lite/.

What this plugin does

• Converts one high-level goal into a trackable delivery flow (Planner -> Developer -> Reviewer).
• Uses state-driven orchestration so work continues predictably instead of stalling in chat context.
• Automatically alternates orchestration sides (Left <-> Right) while preserving progress.
• Recovers from interrupted sessions by resuming in_progress work first.
• Handles failed task attempts with bounded retries and explicit unfinished reasons.
• Stores all runtime state in plain JSON files inside .dreamteam-lite/:
  • .dreamteam-lite/goal.json — immutable project goal.
  • .dreamteam-lite/tasks.json — list of tasks created by the Planner and updated by the orchestrators.
  • .dreamteam-lite/state.json — active orchestrator, batch routing markers, and run status.
• Keeps state human-readable and auditable, so you can inspect and control the workflow at any time.

There is no external engine here: all orchestration lives inside Cursor agents and JSON.

Even though this is a compact demo, it is built to be practically useful end-to-end: small tasks can be completed without babysitting, and the workflow keeps its shape under typical Cursor model limits.

Relationship to the main DreamTeam project

• The original DreamTeam engine and architecture live at
  https://github.com/budagov-lab/DreamTeam.
• DreamTeam Lite:
  • reuses the high-level pattern (Dispatcher + Left/Right, planning vs execution),
  • but removes the SQLite task DAG, advanced maintenance agents, analytics dashboard, and CLI.
  • is designed as a practical learning project and a starting point for experiments.

If you need the full autonomous development stack with large task graphs, dashboards, and advanced runtime tooling, use the main DreamTeam project.
If you want a lightweight, Cursor-native version of the orchestration pattern, DreamTeam Lite is built for that.

Usage (conceptual)

• Primary entrypoints after plugin install:
  • /start-dreamteam <FULL GOAL TEXT> to initialize goal/state and trigger planning.
  • /run-dreamteam to continue execution batches until completion.
• Project-local behavior:
  • Runtime files are stored under .dreamteam-lite/ in the current workspace only.
  • At startup, the Left Orchestrator creates .dreamteam-lite/ if it does not exist.
  • Task state is never shared between different repositories/projects.
• Dispatcher core contract remains /start and /run internally:
  • Goal capture for start is still all text immediately after /start (verbatim).
  • Skills apply the same contract without spawning a separate dispatcher subagent.
• Current execution details:
  • Exactly one task is processed per orchestrator call.
  • If Reviewer returns CRITICAL, the same task is retried up to 2 additional times (3 total attempts).
  • If still not approved, task stays needs_changes with unfinished = true and unfinished_reason.
  • On resume, tasks with status = "in_progress" are prioritized first.
  • If .dreamteam-lite/state.json or .dreamteam-lite/goal.json is malformed (e.g., concatenated JSON objects), agents normalize it to one valid object and continue.

All runtime instructions and UI/console messages must be in English. The documentation here is also in English so that the plugin can be shared easily.

Autonomy Proof (Snake from zero)

• Goal: create a simple Snake game for playing in browser
• Prompt: one user prompt to start the whole workflow
• Execution: AUTO under Cursor free-tier limits
• Result: a complete playable Snake game shipped from scratch (no manual babysitting)
• Runtime: ~37 minutes
• Evidence: generated .dreamteam-lite/state.json ended with run_status = "all_complete" and .dreamteam-lite/tasks.json ended with status = "done" for all tasks

Video (full run):