aurora-actr 0.6.4

AURORA: Adaptive Unified Reasoning and Orchestration Architecture with MCP Integration

Aurora

Memory-First Planning & Multi-Agent Orchestration Framework

Version 0.6.3 | PyPI | Commands | Docs

Aurora is a local-first development automation framework that uses ACT-R memory, SOAR decomposition, and multi-agent orchestration to systematically break down goals and coordinate specialized agents—all without requiring API access.

Built on planning principles adapted from OpenSpec.

---

What Aurora Actually Is

Aurora provides three core capabilities:

Memory (ACT-R)

Intelligent code indexing that learns from usage patterns. Your codebase's memory works like human memory—frequently accessed code stays "hot", rarely used code fades.

Technology:

• ACT-R activation scores (chunks strengthen/weaken with use)
• BM25 keyword search (fast, reliable, local)
• Git-aware indexing (respects .gitignore, tracks changes)
• Tree-sitter parsing (understands code structure)
• Optional semantic search (if you want it)

Use case: aur mem index . ->aur mem search "authentication"

Planning (SOAR + OpenSpec)

Systematic goal decomposition with agent matching and gap detection. Breaks high-level goals into actionable subgoals with automatic agent assignment.

Technology:

• SOAR 9-phase pipeline (assess ->retrieve ->decompose ->verify ->route)
• Agent capability matching with LLM fallback
• Gap detection (identifies missing agent capabilities early)
• Memory-aware (uses indexed code to inform planning)
• OpenSpec-inspired workflow (goals ->PRD ->tasks)

Use case: aur goals "Add OAuth2 auth" ->/plan ->/implement

Orchestration (Multi-Agent)

CLI-agnostic agent execution with parallel/sequential coordination. Routes tasks to specialized agents without API lock-in.

Technology:

• Works with 20+ CLI tools (claude, cursor, aider, etc.)
• Subprocess-based execution (local-first)
• Dependency-aware scheduling (parallel + sequential)
• Agent specialization (right tool for right job)

Use case: /implement executes tasks sequentially, aur spawn tasks.md executes in parallel

---

What Aurora Is NOT

Not deep reasoning - Aurora uses structured decomposition, not chain-of-thought reasoning like o1 Not API-dependent - Core functionality works locally with any CLI tool Not magical AI - It's systematic orchestration, not AGI Not a replacement for thinking - It's a framework for organizing development work

---

Quick Start

# Install
pip install aurora-actr

# Initialize project
cd your-project/
aur init

# Index your codebase
aur mem index .

# Search indexed memory
aur mem search "authentication logic"

# Answer complex questions with SOAR reasoning
aur soar "How does the payment flow work?"

# Planning flow: Goal ->PRD ->Tasks ->Implementation
aur goals "Add Stripe payment processing" \
  --context src/checkout/ \
  --verbose

# Output: .aurora/plans/0001-add-stripe-payment/goals.json
# Contains: 5 subgoals, agent assignments, memory context, gaps

# Navigate to plan directory
cd .aurora/plans/0001-add-stripe-payment/

# Generate PRD and tasks (in Claude Code/Cursor/etc.)
/plan

# Execute tasks sequentially with specialized agents
/implement

# Alternative: Execute tasks in parallel (faster but less controlled)
aur spawn tasks.md --verbose

# Health check
aur doctor

---

The Planning Flow

+---------------+     +---------------+     +---------------+
|  aur goals    | --> |  /plan        | --> |  /implement   |
|               |     |               |     |               |
| SOAR          |     | OpenSpec      |     | Sequential    |
| decomposes    |     | generates     |     | execution     |
| goal into     |     | PRD +         |     | with          |
| subgoals      |     | tasks.md      |     | specialized   |
| with agents   |     |               |     | agents        |
+---------------+     +---------------+     +---------------+
      |                     |                      |
      v                     v                      v
  goals.json           PRD + tasks            Implemented
  (structured)         (detailed)             (tested)

Alternative parallel execution:
                                          +---------------+
                                          |  aur spawn    |
                                          |               |
                                          | Parallel      |
                                          | execution     |
                                          | (faster)      |
                                          +---------------+

Primary workflow:

• aur goals ->SOAR decomposition with agent matching
• /plan ->Generate PRD and tasks (Claude Code skill)
• /implement ->Execute tasks sequentially, one by one

Alternative:

• aur spawn tasks.md ->Execute tasks in parallel (faster, less control)

---

Real Example

# 1. Index codebase
$ aur mem index .
[OK] Indexed 2,431 chunks from 342 files

# 2. Decompose goal with SOAR
$ aur goals "Add Stripe payment processing" \
    --context src/checkout/ \
    --verbose

Using tool: claude (model: sonnet)

📋 Decomposing goal into subgoals...
   Goal: Add Stripe payment processing

Memory search found 8 relevant files:
  - src/checkout/cart.py (0.92)
  - src/orders/models.py (0.87)
  - src/api/payments.py (0.81)

🤖 Agent matching results:
   [OK] sg-1: Set up Stripe SDK (@full-stack-dev, 0.89)
   [OK] sg-2: Create payment endpoints (@full-stack-dev, 0.91)
   [OK] sg-3: Add webhook handlers (@full-stack-dev, 0.85)
   [OK] sg-4: Implement payment UI (@ux-expert, 0.78)
   [WARN] sg-5: PCI compliance (@security-engineer, NOT FOUND)

Agent gaps detected:
  - Missing @security-engineer for sg-5
  - Suggested capabilities: ["PCI DSS", "security audit"]
  - Fallback: @full-stack-dev (review required)

[OK] Goals saved to .aurora/plans/0001-add-stripe-payment/goals.json

Next steps:
1. Review goals:   cat .aurora/plans/0001-add-stripe-payment/goals.json
2. Generate PRD:   cd .aurora/plans/0001-add-stripe-payment && /plan
3. Implement:      /implement (sequential) or aur spawn tasks.md (parallel)

# 3. Navigate and generate PRD
$ cd .aurora/plans/0001-add-stripe-payment/
$ /plan  # In Claude Code

[OK] Generated prd.md (1,234 lines)
[OK] Generated tasks.md (24 tasks)

# 4. Execute tasks sequentially (recommended)
$ /implement

Task 1/24: Set up Stripe SDK... COMPLETE [OK]
Task 2/24: Create payment models... COMPLETE [OK]
Task 3/24: Add API endpoints... COMPLETE [OK]
...
Task 24/24: Update documentation... COMPLETE [OK]

All tasks complete! [OK]

# Alternative: Execute in parallel (faster)
$ aur spawn tasks.md --verbose

Spawning 5 tasks across 3 agents:
  [@full-stack-dev] Task 1.0: Set up Stripe SDK... COMPLETE (45s)
  [@full-stack-dev] Task 2.0: Create endpoints... COMPLETE (67s)
  [@ux-expert] Task 4.0: Payment UI... COMPLETE (89s)
  ...

All tasks complete! [OK]

---

Key Features

Memory System

• ACT-R activation - Code chunks strengthen/weaken based on usage
• Hybrid retrieval - BM25 + activation + optional semantic
• Git-aware - Respects .gitignore, tracks file changes
• Multi-type storage - Code, knowledge base, reasoning patterns
• Sub-500ms search - Fast retrieval on 10K+ chunks

Planning Workflow (OpenSpec-Inspired)

• aur goals - SOAR decomposition with agent matching
• /plan skill - Generate PRD and tasks from goals.json
• /implement - Sequential task execution with specialized agents
• Gap detection - Identifies missing agent capabilities early
• Memory-aware - Uses indexed code to inform planning

Agent Orchestration

• CLI-agnostic - Works with claude, cursor, aider, and 20+ tools
• Local execution - No cloud APIs required for core features
• Sequential + parallel - /implement (careful) or aur spawn (fast)
• Specialized agents - @full-stack-dev, @ux-expert, @qa-architect, etc.

Configuration

• Multi-tier resolution - CLI flag ->env var ->project config ->global config ->default
• Tool/model selection - Per-command configuration
• Project-local - .aurora/config.json overrides global settings

---

Installation

Standard (Lightweight)

pip install aurora-actr  # ~520KB, BM25 + activation only

With Semantic Search (Optional)

pip install aurora-actr[ml]  # +1.9GB, adds sentence-transformers

Note: The ML package uses all-MiniLM-L6-v2 by default. To use different embedding models (OpenAI, Cohere, custom models), see the ML Models Guide.

Development

git clone https://github.com/amrhas82/aurora.git
cd aurora
./install.sh

---

What Gets Indexed

Aurora indexes three types of chunks:

• code - Functions, classes, methods (tree-sitter AST parsing)
• kb - Markdown documentation (README.md, docs/, PRDs)
• soar - Reasoning patterns (auto-saved from aur soar queries)

Default exclusions: .git/, venv/, node_modules/, tasks/, CHANGELOG.md, LICENSE*, build/, dist/

Custom exclusions: Create .auroraignore (gitignore-style patterns):

# .auroraignore example
tests/**
docs/archive/**
*.tmp

---

Retrieval Strategy

Hybrid scoring (default, no ML required):

• 40% BM25 keyword matching
• 30% ACT-R activation (usage frequency + recency)
• 30% Git signals (modification patterns)

With ML option ([ml]):

• 30% BM25 keyword matching
• 40% Semantic similarity (sentence-transformers)
• 30% ACT-R activation

Speed: Sub-500ms on 10K+ chunks.

Custom models: See ML Models Guide for using OpenAI, Cohere, or custom embedding models.

---

Commands

Command	Purpose	Example
aur init	Initialize project	aur init
aur mem index	Index codebase	aur mem index .
aur mem search	Search memory	aur mem search "auth"
aur soar	Answer questions	aur soar "How does auth work?"
aur goals	Decompose goals	aur goals "Add feature"
aur spawn	Execute tasks (parallel)	aur spawn tasks.md
aur agents list	List agents	aur agents list
aur doctor	Health check	aur doctor

Skills (in Claude Code):

• /plan - Generate PRD and tasks from goals.json
• /implement - Execute tasks sequentially

Full Command Reference →

---

Architecture

SOAR Pipeline (9 Phases)

Two execution modes:

1. Query Mode (aur soar) - Answer questions about code

• Uses all 9 phases to gather info and synthesize answers
• Executes research agents to collect data
• Returns natural language answer with citations

2. Goals Mode (aur goals) - Decompose goals for planning

• Uses phases 1-5, 8-9 (skips execution/synthesis)
• Matches subgoals to agents, detects gaps
• Returns structured goals.json for /plan skill

9 Phases:

1. Assess - Determine complexity (keyword + optional LLM)
2. Retrieve - Get relevant context from ACT-R memory
3. Decompose - Break goal into subgoals with agent assignments
4. Verify - Validate decomposition (self or adversarial)
5. Route - Match agents, detect capability gaps
6. Collect - Execute agents (query mode only)
7. Synthesize - Combine outputs (query mode only)
8. Record - Cache successful patterns
9. Respond - Format output (answer or goals.json)

OpenSpec-Inspired Planning

Aurora's planning workflow is inspired by and adapted from OpenSpec:

Core workflow:

• goals.json - Structured goal representation with subgoals
• PRD generation - Detailed product requirements from goals
• Task breakdown - Actionable tasks with agent assignments
• Implementation tracking - Sequential execution with validation

Aurora's extensions:

• ACT-R memory integration for context-aware planning
• SOAR decomposition for systematic goal breakdown
• Agent capability matching with gap detection
• CLI-agnostic multi-agent orchestration

---

Configuration

Global Config (~/.aurora/config.json)

{
  "goals": {
    "default_tool": "claude",
    "default_model": "sonnet"
  },
  "memory": {
    "index_on_save": true
  },
  "logging": {
    "level": "INFO"
  }
}

Project Config (.aurora/config.json)

{
  "goals": {
    "default_tool": "cursor",
    "default_model": "opus"
  }
}

Environment Variables

export ANTHROPIC_API_KEY=sk-ant-...
export AURORA_GOALS_TOOL=claude
export AURORA_GOALS_MODEL=sonnet
export AURORA_LOGGING_LEVEL=DEBUG

Resolution order: CLI flag ->env var ->project config ->global config ->default

Configuration Reference →

---

Documentation

• Commands Reference - Complete CLI command guide
• Tools Guide - Comprehensive tooling ecosystem
• Configuration Reference - All settings and environment variables
• SOAR Architecture - Technical pipeline details
• Planning Flow - End-to-end workflow guide
• Goals Command - Full goals command reference
• ML Models Guide - Custom embedding model configuration
• Migration Guide - Migrating from MCP tools

---

Design Principles

1. Memory-First - ACT-R activation as cognitive foundation
2. Local Execution - No cloud APIs required for core features
3. CLI-Agnostic - Works with any tool, no vendor lock-in
4. Systematic Over Magical - Structured pipelines, not black boxes
5. Honest Capabilities - We decompose and orchestrate, we don't "reason deeply"
6. OpenSpec-Inspired - Proven planning workflow patterns

---

Requirements

• Python 3.10+
• Git (for git-aware indexing)
• One or more CLI tools: claude, cursor, aider, etc.
• Optional: Anthropic API key (for semantic search with ML package)

Disk space:

• Base install: ~520KB
• With ML features: ~1.9GB (PyTorch + sentence-transformers)

---

Credits

• OpenSpec - Planning and implementation workflow patterns
• ACT-R - Cognitive architecture for memory modeling
• SOAR - Cognitive architecture for goal decomposition
• Tree-sitter - Code parsing
• Anthropic Claude - LLM capabilities (when using API)

---

License

MIT License - See LICENSE for details.

---

Contributing

See CONTRIBUTING.md for development setup and guidelines.