claude-code

name	claude-code
description	Delegate coding tasks to Claude Code from Mind Flayer. Use for feature work, refactoring, code review, and iterative coding when the claude CLI is installed.
metadata	{"icon":"assets/icon.svg","requires":{"bins":["claude"]}}

Claude Code Delegation

Use this skill when the user wants Mind Flayer to delegate coding work to Claude Code, Anthropic's autonomous coding CLI.

Prerequisites

• Install: npm install -g @anthropic-ai/claude-code
• Authenticate once with claude, claude auth login --console, or ANTHROPIC_API_KEY
• Check installation with claude --version
• Check auth with claude auth status

Tools

Use Mind Flayer's controlled agent session tools:

• agentSessionStart starts a Claude Code process.
• agentSessionRead reads output from a background non-interactive job.
• agentSessionStop stops a running background job.

These tools only build supported Claude Code commands. Do not try to run arbitrary shell commands through them. Mind Flayer chat does not support Claude Code interactive TUI sessions; use print.

Preferred Mode: Print

For one-shot work, use agentSessionStart with:

{
  "agent": "claude-code",
  "mode": "print",
  "cwd": "/absolute/path/to/project",
  "prompt": "Fix the failing auth tests and explain the changes.",
  "runMode": "foreground",
  "permissionPreset": "workspace-write",
  "timeoutSeconds": 300
}

Use print mode for:

• Focused bug fixes
• Small features
• Refactors with clear scope
• Code review summaries
• Running a bounded investigation

Prefer permissionPreset: "read-only" for reviews and analysis. Use workspace-write only when edits are intended.

Background Jobs

For long-running non-interactive tasks, set runMode to background, then poll with agentSessionRead. Stop abandoned jobs with agentSessionStop.

Safety

• Always set cwd to the exact project directory.
• Do not use the Mind Flayer shared workspace root as a coding-agent project directory.
• Do not use permission bypass modes unless the user explicitly asks and accepts the risk.
• Do not delegate secrets handling unless the user explicitly requested it.
• Summarize what Claude Code did, including changed files and tests run.

Useful Claude Code Patterns

• Review only: use permissionPreset: "read-only".
• Plan only: use permissionPreset: "plan".
• Write code: use permissionPreset: "workspace-write".
• Add extra repo directories with extraAllowedDirs only when the task needs them.

After Delegation

When a session completes:

1. Read the final output with agentSessionRead.
2. Inspect local changes yourself before reporting success.
3. Run the relevant project checks.
4. Report the outcome and any residual risk to the user.