@evolver/adapter-claude-code

Claude Code CLI executor adapter for the Evolver framework.

Installation

pnpm add @evolver/adapter-claude-code

Usage

import { ClaudeCodeExecutor, ResultParser } from "@evolver/adapter-claude-code";

const executor = new ClaudeCodeExecutor({
  name:        "claude-code",
  command:     "claude",
  skillsPath:  "./skills",
  skillFormat: "markdown",
  timeout:     60_000,
  concurrency: 2,
});

// Execute a program against tasks (implements @evolver/core Executor interface)
const results = await executor.run(program, tasks);

API

ClaudeCodeExecutor

Executes tasks via the Claude Code CLI (--print --output-format json --skill-path). Deploys skills as .md files to a temp directory, runs tasks in batches with configurable concurrency, and scores results.

new ClaudeCodeExecutor(config: AdapterConfig)
executor.run(program: Program, tasks: Task[]): Promise<ExecutionResult[]>

Security: uses execFile (no shell invocation) to prevent injection. All arguments are passed as arrays.

Built-in scorers: exact-match (default) and fuzzy.

ResultParser

Static parser for Claude Code CLI stdout/stderr output.

ResultParser.parse(taskId: string, stdout: string, stderr: string, durationMs: number): ExecutionResult

Handles JSON and raw text output. Extracts tokenUsage and promotes result field when present in JSON responses.

Configuration

Uses AdapterConfig from @evolver/core:

| Option | Type | Description | |--------|------|-------------| | name | string | Adapter identifier | | command | string | CLI command (e.g. "claude") | | skillsPath | string | Path to skills directory | | skillFormat | "markdown" \| "json" \| "yaml" | Skill file format | | timeout | number | Per-task timeout in milliseconds | | concurrency | number | Max parallel task executions |

Testing

pnpm test