adelie-ai

v0.3.9

Published

16 days ago

Adelie — Self-Communicating Autonomous AI Loop CLI

0High
0Medium
0Low

bunhine0452

ai autonomous ollama gemini cli agent llm

Overview

Adelie is an autonomous AI orchestrator that plans, codes, reviews, tests, deploys, and evolves software projects through a coordinated multi-agent loop. It ships as a single CLI (npm install -g adelie-ai) and requires only an LLM provider — no cloud backend, no account.

    (o_    Adelie v0.3.5
    //\    gemini · gemini-2.5-pro
    V_/_   Phase: mid_1 | 🛡️3 📡🟢 🧠12/5

Unlike simple code-generation tools, Adelie provides a Harness — a structural framework that constrains, monitors, and self-corrects the AI at every step:

🔧 Dynamic Pipeline — The AI reconfigures its own execution pipeline based on project needs
🛡️ Policy Engine — Declarative constraints block unsafe code before it reaches your project
🧠 Selective Memory — Phase-aware KB filtering prevents context derailment and hallucination
📡 Production Bridge — Live CI/CD and error monitoring feeds failures back into the AI loop
⛔ Human Intercept — Stop the AI instantly at any moment from CLI or Dashboard

What's New in v0.3.0

v0.3.0 is a major architectural evolution. The AI no longer just executes a fixed pipeline — it understands, adapts, and self-corrects its own operational structure.

🔧 Meta Harness — Dynamic Pipeline (v0.2.16)

The static 6-phase pipeline is now a dynamic JSON state machine. The Expert AI can analyze a project and restructure the execution flow at runtime.

# Expert AI decides: "This project needs a security audit phase"
# → MODIFY_HARNESS action creates solidity_auditor agent + Security Audit phase
# → Pipeline automatically restructured: mid → security_audit → mid_1

HarnessManager — Loads/saves/validates JSON-based pipeline configurations
DynamicAgent — Runtime-configurable agents with 3-tier permissions (observer → analyst → operator)
Snapshot Rollback — Every modification creates a backup; failed changes auto-revert

🛡️ Policy Engine — Declarative Constraints (v0.2.17)

Define project-specific rules in .adelie/constraints.yaml that the AI cannot violate:

rules:
  - name: orm-only
    type: ast
    description: "DB queries must use ORM"
    check: no_raw_sql

  - name: timeout-required
    type: pattern
    pattern: 'requests\.(get|post|put|delete)\('
    negative_pattern: 'timeout='
    description: "HTTP calls must include timeout"
    severity: block

AST Analysis — Python AST checks for eval(), exec(), wildcard imports, missing docstrings
Pattern Matching — Regex rules with negative pattern support for false positive suppression
PolicyGate — Blocks staging → project promotion on violation; forces Coder retry

🧠 Memory Harness — Selective Forgetting (v0.2.18)

Prevents context derailment by controlling what the AI can and cannot see:

Phase Scope Filter — KB files tagged with phase_scope are only visible during designated phases
Archive Manager — Resolved errors and completed-phase docs auto-archived after 3 cycles
Summary Tree — archive/summaries.md preserves 1-2 line summaries for historical awareness
Context Budget — Only 5% of context window allocated to archived knowledge

📡 Production Bridge — CI/CD Feedback Loop (v0.2.19)

Connects the AI loop to external production monitoring:

AI deploys code → GitHub Actions fails → Bridge detects CI failure
→ HealthVerdict: CRITICAL → Orchestrator: ERROR state
→ Coder generates hotfix → Deploy again → Bridge confirms: HEALTHY

GitHub Actions Adapter — Polls CI status via REST API or MCP
Sentry Adapter — Monitors error spikes above threshold
Custom MCP Adapter — Auto-discovers production monitoring tools
Graceful Degradation — Missing tokens/config = adapter silently disabled

⛔ Human Intercept & Monitoring Overhaul (v0.2.20)

Complete CLI and Dashboard redesign for real-time visibility and control:

> /intercept AI is going in wrong direction
  ⛔ INTERCEPTED
  Cycle: #12
  State: normal → error
  Use /resume to continue with recovery flow.

> /status
  Status: RUNNING
  Phase: mid_1  |  State: normal  |  Cycle: 5
  🛡️ Policy Engine: 3 rules (2 pattern, 1 ast)
  🧠 Memory Harness: 12 active / 5 archived / 3 scoped
  📡 Production: HEALTHY (github_actions, sentry)
  🔧 Pipeline: 6 phases, 13 agents

New CLI Commands: /intercept, /policy, /health, /memory, /harness

Dashboard Panels: Intercept button, Production Health, Policy Engine, Memory Harness, Pipeline Visualizer

See CHANGELOG.md for complete release history.

Quick Start

Prerequisites

| Requirement | Version | |:--|:--| | Python | 3.10+ | | Node.js | 16+ | | LLM | Gemini API key or Ollama instance |

Install

npm (recommended)

npm install -g adelie-ai

curl (macOS / Linux)

curl -fsSL https://raw.githubusercontent.com/Ade1ie/adelie/main/install.sh | bash

PowerShell (Windows)

irm https://raw.githubusercontent.com/Ade1ie/adelie/main/install.ps1 | iex

Homebrew (macOS / Linux)

brew tap Ade1ie/tap
brew install adelie

From source

git clone https://github.com/Ade1ie/adelie.git
cd adelie
pip install -r requirements.txt
python -c "from adelie.cli import main; main()"

Update

# Built-in update checker
adelie --update

# or manual
npm install -g adelie-ai@latest

# Homebrew
brew upgrade adelie

# Check current version
adelie --version

Configure

cd your-project/
adelie init

# Gemini
adelie config --provider gemini --api-key YOUR_KEY

# or Ollama (local, free)
adelie config --provider ollama --model gemma3:12b

Run

# Continuous autonomous loop
adelie run --goal "Build a REST API for task management"

# Single cycle
adelie run --once --goal "Analyze and document the codebase"

# Resume a saved workspace
adelie run ws 1

The real-time dashboard opens automatically at http://localhost:5042.

How It Works

Agents

Adaptive Lifecycle

initial → mid → mid_1 → mid_2 → late → evolve
                  ↑
          Expert AI can insert custom
          phases here at runtime

Each phase transition is gated by quality metrics — KB file count, test pass rate, review scores, stability indicators. The Expert AI votes on phase transitions; the Harness enforces the gates.

Harness System

The Harness is what makes Adelie different from a simple AI code generator. It's a structural framework that constrains and protects the AI:

┌─────────────────────────────────────────────────┐
│                  HARNESS LAYER                   │
│                                                  │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐        │
│  │ Policy   │ │ Memory   │ │Production│        │
│  │ Engine   │ │ Harness  │ │ Bridge   │        │
│  │          │ │          │ │          │        │
│  │constraints│ │phase scope│ │CI/CD     │        │
│  │AST check │ │archiving │ │Sentry    │        │
│  │pattern   │ │summaries │ │GitHub    │        │
│  └────┬─────┘ └────┬─────┘ └────┬─────┘        │
│       │            │            │               │
│  ┌────▼────────────▼────────────▼─────┐         │
│  │         ORCHESTRATOR               │         │
│  │  cycle start → agents → promote    │         │
│  └────────────────────────────────────┘         │
│                                                  │
│  ┌──────────┐ ┌──────────┐                      │
│  │ Harness  │ │ Human    │                      │
│  │ Manager  │ │ Intercept│                      │
│  │          │ │          │                      │
│  │dynamic   │ │/intercept│                      │
│  │pipeline  │ │dashboard │                      │
│  │rollback  │ │⛔ button │                      │
│  └──────────┘ └──────────┘                      │
└─────────────────────────────────────────────────┘

Security

Adelie enforces multiple security layers:

Policy Engine — Declarative constraints block AST violations, forbidden patterns, and file limits
Shell injection prevention — BLOCKED_CHARS filter blocks &, >, |, ;, backticks
Path traversal protection — Path.resolve() verification ensures files stay within staging
Staging isolation — Code is written to .adelie/staging/ first, verified, then promoted
Thread-safe operations — _usage_lock and _staging_lock prevent race conditions
Whitelisted commands — Runner and Tester only execute pre-approved command patterns

Architecture

Dashboard

Adelie serves a real-time monitoring UI at http://localhost:5042 (auto-starts with adelie run).

Agent grid — live status of all agents (idle / running / done / error)
⛔ Intercept button — emergency stop from the browser
Production Health — real-time verdict badge (🟢 HEALTHY / 🟡 DEGRADED / 🔴 CRITICAL)
Policy Engine — active rule count and type breakdown
Memory Harness — active/archived/scoped file counters
Pipeline Visualizer — horizontal phase flow with active/completed highlighting
Log stream — real-time SSE-powered log feed with category filtering
Cycle metrics — tokens, LLM calls, files written, test results, review scores
Cycle history chart — last 30 cycles at a glance

Built with zero external dependencies — Python http.server + SSE + embedded HTML/JS.

CLI

adelie --version                  # Show version
adelie --update                   # Check for updates
adelie help                       # Full command reference

Workspace

adelie init [dir]                 # Initialize .adelie workspace
adelie ws                         # List all workspaces
adelie ws remove <N>              # Remove workspace
adelie scan                       # Scan existing codebase → KB

Execution

adelie run --goal "…"             # Start continuous loop
adelie run --once --goal "…"      # Single cycle
adelie run ws <N>                 # Resume workspace #N

Interactive Commands (REPL)

Configuration

adelie config                     # Show current config
adelie config --provider ollama   # Switch LLM provider
adelie config --model gemma3:12b  # Set model
adelie config --api-key KEY       # Set Gemini API key
adelie config --plan-mode true    # Enable Plan Mode (human approval)

Monitoring

adelie status                     # System health & provider status
adelie inform                     # AI-generated project report
adelie phase                      # Show current phase
adelie metrics                    # Cycle metrics & history
adelie metrics --agents           # Per-agent token usage

Knowledge Base & Project

adelie kb                         # KB file counts by category
adelie kb --clear-errors          # Clear error files
adelie goal                       # Show project goal
adelie goal set "…"               # Set project goal
adelie feedback "message"         # Inject feedback into AI loop
adelie research "topic"           # Web search → KB
adelie spec load <file>           # Load spec (MD/PDF/DOCX) into KB

Configuration

Environment (`.adelie/.env`)

Policy Engine (`.adelie/constraints.yaml`)

rules:
  - name: no-eval
    type: pattern
    pattern: '\beval\s*\('
    description: "Block eval() calls"
    severity: block

  - name: docstrings
    type: ast
    check: missing_docstring
    description: "Public functions must have docstrings"
    severity: warn

Production Bridge

# .adelie/.env
PRODUCTION_BRIDGE_ENABLED=true

# GitHub Actions
GITHUB_TOKEN=ghp_xxxxx

# Sentry
SENTRY_AUTH_TOKEN=sntrys_xxxxx
SENTRY_ORG=my-org
SENTRY_PROJECT=my-project

Or add MCP servers to .adelie/mcp.json for MCP-based monitoring.

Platform Features

Testing

python -m pytest tests/ -v    # 748 tests

Project Structure

adelie/
├── orchestrator.py          # Main loop — state machine + harness integration
├── harness_manager.py       # Dynamic pipeline configuration + rollback
├── policy_engine.py         # Declarative constraint enforcement
├── memory_harness.py        # Selective forgetting + phase-aware KB
├── production_bridge.py     # CI/CD + monitoring feedback loop
├── commands/                # CLI command modules
├── cli.py                   # CLI entry point + argparse routing
├── config.py                # Configuration & env loading
├── llm_client.py            # LLM abstraction (Gemini + Ollama + fallback)
├── interactive.py           # REPL + dashboard + intercept commands
├── dashboard.py             # Real-time web server (HTTP + SSE)
├── dashboard_html.py        # Embedded dashboard UI template
├── agents/                  # 12+ specialized AI agents
│   ├── writer_ai.py         #   Knowledge Base curator
│   ├── expert_ai.py         #   Strategic decision maker
│   ├── coder_ai.py          #   Code generator
│   ├── coder_manager.py     #   Layer dispatch & retry
│   ├── reviewer_ai.py       #   Quality reviewer
│   ├── tester_ai.py         #   Test runner
│   ├── runner_ai.py         #   Build & deploy
│   ├── monitor_ai.py        #   Health monitor
│   ├── analyst_ai.py        #   Trend analyzer
│   ├── research_ai.py       #   Web researcher
│   ├── inform_ai.py         #   Status report generator
│   ├── scanner_ai.py        #   Initial codebase scanner
│   └── dynamic_agent.py     #   Runtime-created agents
├── utils/
│   └── ast_checker.py       #   AST-based static analysis
├── kb/                      # Knowledge Base (retriever + embeddings)
├── channels/                # Multichannel providers (Discord, Slack)
├── a2a/                     # Agent-to-Agent protocol
├── checkpoint.py            # Snapshot & rollback
├── sandbox.py               # Docker/Seatbelt isolation
├── gateway.py               # REST API gateway
├── skill_manager.py         # Skill registry
├── plan_mode.py             # Plan Mode (human approval)
├── loop_detector.py         # Stuck-pattern detection
├── scheduler.py             # Per-agent scheduling
├── phases.py                # Lifecycle phase definitions (compat shim)
├── hooks.py                 # Event-driven plugin system
└── process_supervisor.py    # Subprocess management

Contributing

git clone https://github.com/Ade1ie/adelie.git
cd adelie
pip install -r requirements.txt
python -m pytest tests/ -v   # Ensure all tests pass

Fork → branch → implement → test → PR
Follow existing code style and patterns
Add tests for new features

License

MIT

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

Overview

What's New in v0.3.0

🔧 Meta Harness — Dynamic Pipeline (v0.2.16)

🛡️ Policy Engine — Declarative Constraints (v0.2.17)

🧠 Memory Harness — Selective Forgetting (v0.2.18)

📡 Production Bridge — CI/CD Feedback Loop (v0.2.19)

⛔ Human Intercept & Monitoring Overhaul (v0.2.20)

Quick Start

Prerequisites

Install

npm (recommended)

curl (macOS / Linux)

PowerShell (Windows)

Homebrew (macOS / Linux)

From source

Update

Configure

Run

How It Works

Agents

Adaptive Lifecycle

Harness System

Security

Architecture

Dashboard

CLI

Workspace

Execution

Interactive Commands (REPL)

Configuration

Monitoring

Knowledge Base & Project

Configuration

Environment (.adelie/.env)

Policy Engine (.adelie/constraints.yaml)

Production Bridge

Platform Features

Testing

Project Structure

Contributing

Links

License

Environment (`.adelie/.env`)

Policy Engine (`.adelie/constraints.yaml`)