@evolver/cli

Command-line interface for the Evolver skill evolution framework.

Installation

pnpm add -g @evolver/cli

Usage

evolver evolve -- Run the evolution loop

evolver evolve \
  --task-dir ./tasks \
  --adapter claude-code \
  --max-iterations 10 \
  --frontier-capacity 3 \
  --budget-limit 5.0

evolver status -- Show last run status

evolver status

Reads .evolver/state.json and prints last run timestamp, iterations, cost, best program score, frontier size, and discovered skills.

evolver skills list -- List discovered skills

evolver skills list --skills-dir ./skills

evolver skills export -- Export skills to another format

evolver skills export --format cursor --output .cursorrules

evolver skills test -- Cross-model transfer test

evolver skills test \
  --cross-model \
  --task-dir ./tasks \
  --source claude-code \
  --target cursor,codex

API

Commands

| Command | Description | |---------|-------------| | evolve | Run the skill evolution loop | | status | Show current evolution status from .evolver/state.json | | skills list | List discovered skills in the skills directory | | skills export | Export skills to another agent format (e.g. Cursor .cursorrules) | | skills test | Run cross-model skill transfer tests |

Modules

loadConfig(taskDir: string): TaskConfig

Parses config.yaml from the task directory.

loadTasks(taskDir: string, subDir: string, defaultScorer?: ScorerType): Task[]

Loads YAML task files from {taskDir}/{subDir}/ into Task[] arrays.

saveState(state: EvolverState): void / loadState(): EvolverState | null

Persists/loads evolution state to/from .evolver/state.json.

Configuration

evolve command options

| Option | Default | Description | |--------|---------|-------------| | --task-dir <path> | (required) | Path to tasks directory | | --skills-dir <path> | ./skills | Output skills directory | | --adapter <name> | claude-code | Executor adapter (claude-code, cursor, codex) | | --proposer-model <model> | claude-sonnet-4-6 | Model for proposal generation | | --builder-model <model> | claude-haiku-4-5 | Model for skill materialization | | --runs <n> | 3 | Runs per evaluation | | --budget-limit <usd> | none | Maximum budget in USD | | --frontier-capacity <n> | 3 | Pareto frontier capacity | | --max-iterations <n> | 10 | Maximum evolution iterations | | --failure-threshold <n> | 0.5 | Score threshold for failure | | --plugin <name> | none | Plugin to load (e.g. memento) | | --memento-url <url> | none | Memento MCP server URL | | --memento-key <key> | none | Memento MCP access key |

Task directory structure

tasks/
  config.yaml          # Optional: default scorer, categories
  train/
    task-001.yaml      # { id, input, expected, category?, scorer? }
    task-002.yaml
  validation/
    task-010.yaml

Testing

pnpm test