Claude Code¶

Anthropic's AI coding assistant with deep reasoning capabilities and a focus on safety.

Overview¶

Claude Code is a powerful AI coding assistant from Anthropic. It excels at:

Complex multi-step reasoning
Thorough code analysis and review
Safe and responsible AI assistance
Understanding context deeply

Installation¶

Install Claude Code from Anthropic:

# Verify installation
which claude
claude --version

Basic Usage¶

# Use Claude with clinvk
clinvk --backend claude "fix the bug in auth.go"
clinvk -b claude "explain this codebase"

Models¶

Model	Description
`claude-opus-4-5-20251101`	Most capable, best for complex tasks
`claude-sonnet-4-20250514`	Balanced performance and speed

Specify a model:

clinvk -b claude -m claude-sonnet-4-20250514 "quick review"

Configuration¶

Configure Claude in ~/.clinvk/config.yaml:

backends:
  claude:
    # Default model
    model: claude-opus-4-5-20251101

    # Tool access (all, or comma-separated list)
    allowed_tools: all

    # Override unified approval mode
    approval_mode: default

    # Override unified sandbox mode
    sandbox_mode: default

    # Enable/disable this backend
    enabled: true

    # Custom system prompt
    system_prompt: ""

    # Extra CLI flags
    extra_flags: []

Environment Variable¶

export CLINVK_CLAUDE_MODEL=claude-sonnet-4-20250514

Approval Modes¶

Claude supports different approval behaviors:

Mode	Description
`default`	Let Claude decide based on action risk
`auto`	Reduce prompts / auto-approve edits (backend-specific)
`none`	Never ask for approval prompts (dangerous)
`always`	Always ask for approval (when supported)

Set via config:

backends:
  claude:
    approval_mode: auto

Or per-command (in tasks/chains):

{
  "backend": "claude",
  "prompt": "refactor the module",
  "approval_mode": "auto"
}

Sandbox Modes¶

Control Claude's file system access:

Note

sandbox_mode is a unified setting. For the claude backend, clinvk does not currently map sandbox_mode to a Claude CLI flag, so it may have no effect.

Mode	Description
`default`	Let Claude decide
`read-only`	Can only read files
`workspace`	Can modify files in project
`full`	Full file system access

Allowed Tools¶

Control which tools Claude can use:

backends:
  claude:
    # All tools
    allowed_tools: all

    # Specific tools only
    allowed_tools: read,write,edit

Session Resume¶

Claude Code stores sessions and supports resuming:

# Resume with clinvk
clinvk resume --last --backend claude
clinvk resume <session-id>

Internally uses Claude's --resume flag.

Extra Flags¶

Pass additional flags to the Claude CLI:

backends:
  claude:
    extra_flags:
      - "--add-dir"
      - "./docs"

Common flags:

Flag	Description
`--add-dir <path>`	Add additional directory to context
`--print`	Print final response in non-interactive mode

Best Practices¶

Use Opus for Complex Tasks

Claude Opus is ideal for multi-step reasoning, code architecture, and thorough reviews.

Leverage Session Continuity

Claude excels at maintaining context across a conversation. Use clinvk -c to continue sessions.

Trust the Defaults

Claude's default approval and sandbox modes are well-tuned for safety while being useful.

Use Cases¶

Code Review¶

clinvk -b claude "review this PR for security issues and code quality"

Complex Refactoring¶

clinvk -b claude "refactor the authentication system to use JWT tokens"

Architecture Analysis¶

clinvk -b claude "analyze this codebase architecture and suggest improvements"

Bug Investigation¶

clinvk -b claude "investigate why the tests are failing in the CI pipeline"

Troubleshooting¶

Backend Not Available¶

# Check if Claude is installed
which claude

# Check clinvk detection
clinvk config show | grep claude

Rate Limits¶

If hitting rate limits, consider:

Using a different model
Spacing out requests
Running in sequential mode for comparisons