AROS Meta Loop

Autonomous meta-cognition engine for the AROS ecosystem. Runs a 6-step self-improvement cycle that monitors agent performance, evaluates meta-goals, and tunes operational policy — with human-in-the-loop guardrails.

Architecture

PERCEIVE → SELF-MODEL UPDATE → CRITIQUE → POLICY REVISION → IDENTITY CHECK → PERSIST
    ↑                                          |                    |            |
    |          Channel A (re-query)  ←─────────┘                    |            |
    |          Channel B (rejection) ←──────────────────────────────┘            |
    |          Channel E (drift restart) ←──────────────────────────────────────┘
    └── Channel G (persist confirmation) ←──────────────────────────────────────┘

Feedback channels A–G provide intra-cycle and inter-cycle loops with bounded iteration guards.

Core Components

Component	Description
MetaLoopEngine	6-step cycle with async lock, abort flag, and phase gates
L1 Collector	Operational metrics from meta_events (tokens, success rate, errors)
L2 Evaluator	Scores 6 meta-goals (G1 Truthful → G6 Self-Know) from 0.0–1.0
L3 Signal Deriver	Cross-validated signals (drift, calibration, recurrence)
PolicyChangeClassifier	AUTO_APPROVE / HUMAN_REVIEW / NEVER permission tiers
CadenceController	Rate limits (4 cycles/hr, 20/day, 2 emergency/hr) with per-step timeouts
StateManager	Atomic TOML config, JSONL evolution log, file-based signal queue

Meta-Goals (G1–G6)

Goal	Metric	Default Threshold
G1 Truthful	1.0 − error_rate	0.80
G2 Efficient	Token efficiency vs 50K baseline	0.70
G3 Reliable	Tasks without retries / total	0.85
G4 Aligned	Tool call success rate	0.90
G5 Ambitious	Task throughput proxy	0.50
G6 Self-Know	Calibration accuracy	0.60

Permission Model

• AUTO_APPROVE: Parameter tuning within ±20%, tightening constraints
• HUMAN_REVIEW: Meta-goal changes, loosening constraints, >20% change
• NEVER: Modifying L1 event log, disabling Critic, removing human-review

API Endpoints

POST /api/meta-loop/trigger          Start a cycle
GET  /api/meta-loop/status           Status + L2 scores
POST /api/meta-loop/signal           Inject signal (low/normal/high/urgent)
GET  /api/meta-loop/evolution-log    Paginated cycle history
POST /api/meta-loop/approve/{id}     Approve pending change
GET  /api/meta-loop/pending-approvals
POST /api/meta-loop/event            Webhook (receives events from mini-claude-bot)
POST /api/meta-loop/nirmana          Activate/deactivate autonomous mode

Cadence Modes

Mode	Interval	Use Case
balanced	4 hours	Normal operation
aggressive	30 minutes	Nirmana /away mode
conservative	8 hours	Low-priority background
frozen	Paused	Manual-only triggers

State Directory (~/.aros/)

~/.aros/
├── meta-cognition.toml    # Cadence, goals (G1–G6)
├── self-model.toml        # Capabilities, calibration
├── policy.toml            # Harness and meta-loop params
├── evolution-log.jsonl    # Append-only cycle audit trail
├── data/
│   └── meta-loop.db       # SQLite (meta_events, meta_iterations)
├── signals/               # File-based signal queue
├── pending-review/        # Changes awaiting human approval
└── state/
    ├── last_commit.json   # Channel G persistence
    └── pending_evals.json # Channel F delayed evaluations

Setup

# Install
make install

# Run tests (115 tests)
make test

# Start service on port 8200
make run

Requires Python 3.12+.

Integration

• mini-claude-bot: MetaLoopBridge pushes events via webhook, injects status into Claude context
• telegram-claude-hero: /metastatus command with G1–G6 progress bars
• MCP tools: 6 proxy tools (status, trigger, signal, approve, etc.)

License

Apache License 2.0 — see LICENSE for details.