Claude Code Engine

The Claude Code engine invokes Anthropic's claude CLI.

Prerequisites

• Install Claude Code: ~/.local/bin/claude or available on PATH
• Authenticate: claude auth login or set ANTHROPIC_API_KEY environment variable

Configuration

# golem.yaml
name: my-bot
engine: claude-code
model: claude-sonnet-4-6   # optional, see below
skipPermissions: true       # default: true

Choosing a Model

Model names are Anthropic model IDs, passed directly as --model to the claude CLI.

List available models:

claude models

Latest models:

Model ID	Alias	Description
claude-opus-4-6	claude-opus-4-6	Most capable, best for complex tasks
claude-sonnet-4-6	claude-sonnet-4-6	Balanced speed and intelligence — recommended
claude-haiku-4-5-20251001	claude-haiku-4-5	Fastest, lightweight

See the full and up-to-date list at Anthropic model documentation.

Override at runtime — pass model to createAssistant():

const bot = createAssistant({ dir: './my-bot', model: 'claude-opus-4-5' })

How It Works

CLI Invocation

claude -p "<prompt>" \
  --output-format stream-json \
  --verbose \
  --dangerously-skip-permissions

Optional flags:

• --resume <sessionId> — resume a previous session
• --model <model> — specify model

The --verbose flag is required for intermediate stream events (tool calls, tool results).

Permission Bypass

skipPermissions defaults to true. When enabled, --dangerously-skip-permissions is passed to the CLI. A one-time warning is emitted to stderr. Set skipPermissions: false in golem.yaml to disable this behavior (the agent will prompt for permission on certain actions).

Skill Injection

Skills are symlinked into .claude/skills/:

my-bot/
├── .claude/
│   └── skills/
│       ├── general -> ../../skills/general
│       └── im-adapter -> ../../skills/im-adapter
├── CLAUDE.md -> AGENTS.md
└── skills/
    ├── general/
    └── im-adapter/

Additionally, CLAUDE.md is created as a symlink to AGENTS.md, allowing Claude Code to read the auto-generated assistant context.

Output Parsing

Claude Code emits clean JSON (no ANSI codes). The parser handles:

• assistant messages — text content blocks and tool_use blocks
• user messages — tool_result blocks
• result messages — final result with costUsd (total_cost_usd) and numTurns (num_turns)

Cost & Turn Tracking

Claude Code is the only engine that provides per-conversation cost and turn count in the done event:

{ type: 'done', sessionId: '...', durationMs: 12345,
  costUsd: 0.042, numTurns: 3 }

Environment

GolemBot deletes CLAUDECODE and CLAUDE_CODE_ENTRYPOINT environment variables before spawning, to allow nested invocations of Claude Code.

Notes

• The CLAUDE.md symlink is the standard way Claude Code discovers project instructions — by pointing it to AGENTS.md, the agent sees the full skill list and conventions on startup
• Session resume failures are automatically handled with a fresh session fallback