evolve

What problem does it solve?

This skill automates a disciplined, evaluator-driven evolution loop so agents can iteratively improve code, prompts, algorithms, or pipelines while preserving safety gates and durable memory. It removes ad-hoc trial-and-error by enforcing a preflight confirmation, explicit evaluator timeouts, and a recorded experiment database so each round is traceable and reproducible.

Core Features & Use Cases

• Preflight gating that collects objective, core score, evaluator command/script, explicit timeout, mutation scope, sampling algorithm, and approval before any mutation or evaluation runs.
• Four-stage loop (learn → design → experiment → analyze) with two durable memory systems: a cognition store for reusable external insights and an experiment database for node lineage, scores, and analyses.
• Deterministic CLI toolset (brief, cognition, db, eval, files, summary) to normalize specs, seed and query cognition, sample and record nodes, run evaluators with enforced timeouts, and manage writable file scopes.
• Pluggable samplers and safety including ucb1/greedy/random/island and custom sampler hooks, serialized database access, best-snapshot persistence, and path guards to prevent unauthorized file writes.
• Use Case: Run reproducible architecture or prompt search where each candidate is evaluated by an automated metric, and lessons are carried forward to avoid wasted retries.

Quick Start

Normalize the run brief, set an explicit evaluator command and timeout, confirm preflight with the user, and then execute iterative rounds anchored on sampled parents using the provided CLI wrappers.