🧠 Ares 2.0: Agent Orchestrator

A production-grade, deterministic multi-agent orchestrator CLI. Ares routes tasks to the best localized executor (Claude, Codex, or Cursor) based on a strategic planning phase powered by local Ollama.

graph TD
    User["User Task"] --> Router["Router (Ares CLI)"]
    Router --> Planner["Ollama Planner (Local qwen3:latest)"]
    Planner --> Router
    Router --> TinyTask["Tiny Task Layer (Diagnostics)"]
    TinyTask --> Router
    Router --> Selector["Model Selector (Config-Driven)"]
    Selector --> Router
    Router --> Quota["Quota Manager"]
    Router --> Logger["Task Logger (UUID)"]
    Router --> Git["Git Manager (Auto-branch)"]
    Router --> Adapter["Engine Adapter (Claude/Codex/Cursor)"]
    Adapter --> Router
    Router --> Verify["Post-Fix Verification"]

• Planning Layer: Uses local Ollama model to classify tasks, assign risk/confidence scores, and decompose tasks into discrete slices.
• Tiny Task Layer (Diagnostics): Offloads raw terminal output parsing and diff summarization to local Ollama, reducing token load on Claude by 60-80%.
• Routing Layer: Deterministic rules in config/ares/models.yml allocate tasks to engines based on type and risk.
• Automated Fix Loop: Detects failures, summarizes them locally, escalates for a fix, and re-verifies automatically.
• Traceability: Every task receives a UUID and is logged in logs/UUID.json.
• Safety: Built-in quota tracking and confidence-based escalation to Claude Opus for high-risk work.

Ares 2.0 includes a professional, interactive dashboard for real-time task management.

bin/ares --tui

• Live Monitoring: Track quotas, task history, and AI reasoning live.
• In-App Config: Change models and Ollama settings without leaving the CLI.
• Universal Fixes: Trigger test, lint, or syntax healing with one keypress.

• TUI Guide: Master the interactive dashboard.
• Self-Healing Guide: How Ares automatically repairs your code.
• Configuration Guide: Tuning model routing and local AI parameters.

gem install ares-runtime

Install dependencies:

bundle install

Run a task:

# Using the installed gem
ares "Task description"

# Or using the local executable
exe/ares "Task description"

To run tests and automatically attempt fixes for failures:

exe/ares "run tests"

• -d, --dry-run: Plan and select model without execution.
• -g, --git: Auto-branch before execution and auto-commit results.

Routed via config/models.yml:

• Architecture: Claude Opus (High Reasoning)
• Refactor: Claude Sonnet (Primary Executor)
• Bulk Patch / Test Gen: Codex (High Speed)
• Interactive Edit: Cursor Agent (Human-in-the-loop)
• Summarization: Claude Haiku (Low Cost)

1. Deterministic One-Hop: No recursive agent loops.
2. Quota Aware: Claude usage is tracked daily via QuotaManager.
3. Workspace Isolation: Execution is pinned to the current directory's context via recursive AGENTS.md discovery.
4. Git Protection: .gitignore automatically excludes logs/, *.gem, and AI-specific session/cache directories (.claude, .cursor, etc.).

• config/models.yml: Define routing logic and confidence thresholds.
• config/workspaces.yml: Register explicit workspace roots.
• config/planner_schema.rb: The strict JSON schema for the Ollama planner.

• Cost Tracker: USD/token cost calculation per engine.
• Parallel Execution: Execute independent task slices concurrently.
• Automatic Diff Chunking: Handle massive diffs by chunking before LLM processing.
• Rate Limiting: Intelligent throttling to prevent mid-workflow blocks.

Author: Shubham (shubhamtaywade82@gmail.com) Repository: github.com/shubhamtaywade82/agent-orchestrator