Documentation menu

Triggers

Self-modification runs in two cases:

  1. Deep sleep: automatically, every 10th consolidation cycle.
  2. Explicit request: the creature calls the request_evolution tool.

Both invoke the same evaluation pipeline.

The Creator

A separate LLM conversation spins up with a “Creator” persona, an evolutionary architect whose job is to evaluate and improve the creature’s cognitive code. This is not the creature talking to itself; it’s a distinct context with its own system prompt focused on code quality, effectiveness, and architectural fitness.

Context

The Creator receives:

  • Recent dream reflections and observations
  • Event history (actions, errors, patterns)
  • Rollback history from .sys/rollbacks.jsonl
  • Previous evaluation logs from .self/creator-log.jsonl

This gives it a picture of how the creature has been performing and what past modifications succeeded or failed.

Evaluation Loop

The Creator gets two tools: bash (scoped to /creature) and done. The loop runs for up to 20 turns:

  1. Read: inspect current source, observations, dream logs
  2. Diagnose: identify inefficiencies, rule violations, poor consolidation patterns, wasteful action spending
  3. Modify: edit source files to address highest-leverage issues
  4. Validate: run TypeScript compilation to catch errors
  5. Commit: git commit the changes
  6. Done: end evaluation with written reasoning

The Creator focuses on highest-leverage changes: small modifications that meaningfully improve how the creature thinks, remembers, or acts.

What Gets Evaluated

  • Effectiveness: is the creature making progress on its purpose?
  • Rule compliance: is it following its own guidelines and constraints?
  • Consolidation quality: are dreams producing useful observations?
  • Action economy: is it spending tool calls wisely or looping?
  • Architecture: are there structural improvements to the cognitive loop?

Safety

  • TypeScript validation: changes must compile before being committed.
  • Git commits: every modification is committed, creating an audit trail.
  • Rollback tracking: .sys/rollbacks.jsonl records failed promotions so the Creator knows what didn’t work.
  • Evaluation logs: .self/creator-log.jsonl stores the Creator’s reasoning for each session.
  • Health gate: after the Creator commits, the supervisor’s normal health check applies: 10 seconds of stability before promotion, automatic rollback on crash.