Claude Code
Claude Code is Anthropic's agentic coding tool that lets you collaborate with Claude directly from your terminal.
Installation
See the official installation guide.
For containerized Dagu runs, see Run Claude Code in a Container.
Base Invocation
claude -p "<prompt>"Common Flags
Any with key other than the reserved keys prompt, stdin, provider, fallback, managed, and acp_permission is passed as a CLI flag. Snake-case keys are normalized to kebab-case.
| Flag | Type | Description |
|---|---|---|
--model | string | Model variant (e.g. sonnet, opus) |
--bare | boolean | Minimal mode: skips hooks, LSP, plugin sync, auto-memory, CLAUDE.md discovery, and keychain reads |
--max-turns | number | Maximum tool-calling iterations |
bare: true skips keychain reads, so a subscription login is not visible to the step and the run fails with Not logged in · Please run /login. Use it only when the credential comes from ANTHROPIC_API_KEY in the step environment.
Example
steps:
- name: review-pr
action: harness.run
with:
provider: claude
prompt: |
Review the current branch for bugs and style issues
model: sonnetGenerated invocation:
claude -p "Review the current branch for bugs and style issues" --model sonnetYOLO Mode
By default, Claude Code asks for approval before running commands. A harness.run step has no terminal to answer those prompts, so run in non-interactive mode:
steps:
- name: auto-refactor
action: harness.run
with:
provider: claude
prompt: |
Refactor the auth module
dangerously_skip_permissions: trueGenerated invocation:
claude -p "Refactor the auth module" --dangerously-skip-permissionsWarning: This skips all safety checks. Only use in isolated or trusted environments.
Push-back
A human task after a Claude Code step lets a reviewer send the work back:
steps:
- id: implement
action: harness.run
with:
provider: claude
prompt: |
Implement the requested feature
model: opus
- id: review
depends: [implement]
action: human.task
with:
prompt: Review the implementation
push_back:
rewind_to: implement
form:
type: object
required: [feedback]
properties:
feedback:
type: stringOn push-back, the prompt is augmented with reviewer feedback and the previous stdout log path. See Agent Loops for repeating the step until a check passes and for review workflows.
See Also
- Claude Code CLI Reference: Complete flag reference
- Permission Modes: How approval and YOLO mode work
- Environment Variables:
CLAUDE_CODE_*environment variables - Common Workflows: Example prompts and patterns
- Settings: Global and project-level configuration

