Skip to content

ilteris/actor-orchestrator

Repository files navigation

Actor-Orchestrator: Agentic Swarm Infrastructure

A "Unix-for-AI" architecture for managing concurrent, autonomous agent workstreams using gemini-cli and zmx.

Swarm Command Center

What is it?

The Actor-Orchestrator is a hierarchical agentic system that transforms a flat TODO.md file into an active project blackboard. It uses an Actor Model pattern to delegate tasks to isolated sub-agents (Workers) that operate in their own persistent terminal sessions, managed by a high-fidelity TUI Command Center.

The Vision: High-Fidelity Autonomy

  • Cognitive Multiplier: Offload high-throughput implementation to parallel agents while maintaining high-level architectural oversight.
  • Physical Isolation: Each worker runs in its own zmx session with unique temporary workspaces, preventing file-system conflicts and logical crosstalk.
  • Observability: A dedicated Swarm Command Center provides real-time "Active Reasoning" previews, event logging (PR detection), and interactive "Jack-In" capabilities.
  • Self-Healing: Automated reconciliation loops detect dead workers, audit logs for completion signatures, and update task states without human intervention.
  • Ghost Protocol: Hermetic execution runners that self-destruct upon completion, leaving the system temporary directories pristine.

The Hierarchy

  1. Meta-Orchestrator (Teddy): Bootstraps the environment and launches the Supervisor.
  2. Supervisor Actor: Monitors the .tasks/ directory and TODO.md blackboard; delegates tasks via a deterministic dispatch engine.
  3. Worker Actors: Distributed engineers that execute specific tasks (Clone -> Branch -> Implement -> Verify -> Commit -> PR).

Command Center Interface

  • Active Workstreams: Displays live reasoning and "thoughts" from running workers with a subtle pulsating heartbeat.
  • Task Ledger: A recency-sorted list of the latest 10 tasks and their current states.
  • System Events: A descending feed of infrastructure events, including automated PR link capture.
  • Interactivity:
    • [S] Stream Log: Focus on a specific worker's full log output.
    • [R] Refresh: Force a data re-sync.
    • [Q] Quit: Graceful, silent exit.

S5 Protocol (Atomic Contribution)

All workers follow a strict Git lifecycle:

  1. Isolation in a dedicated branch: task-<ID>-<slug>.
  2. Implementation and local verification.
  3. Automatic push to remote and Pull Request creation via GitHub CLI (gh).
  4. URL reporting back to the master Command Center.

Installation & Setup

Ensure you have the core dependencies installed:

# Install zmx (The persistent terminal layer)
brew tap neurosnap/tap && brew install zmx

# Install gemini-cli (The agent logic)
brew install gemini-cli

# Pre-authorize tools in ~/.gemini/settings.json
# Ensure "run_shell_command", "write_todos", "read_file", "list_files", "delegate_to_agent" are allowed.

Launching a Swarm

  1. Link the Extension:
    cd ~/Code/actor-orchestrator && gemini extensions link .
  2. Activate Mission Control: Inside any project with a .tasks/ directory or TODO.md:
    ./../actor-orchestrator/commands/swarm-launch

Technical Specs

  • State Engine: .tasks/*.json (Individual task metadata).
  • Communication: Shared swarm.log + synchronous tool returns.
  • Runtime: Python 3.x (Dashboard) + Bash (Runners) + Gemini CLI (Agents).

About

A 'Unix-for-AI' agentic swarm orchestrator using gemini-cli and zmx.

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors