菜单
SDK-based Claude agent with full Task tool access. Use for executing handoff prompts or spawning nested sub-agents.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | super-agent |
| description | SDK-based Claude agent with full Task tool access. Use for executing handoff prompts or spawning nested sub-agents. |
| allowed-tools | Bash |
SDK-based Claude instance with full tool access including Task (can spawn its own sub-agents).
# Execute a prompt file
PROMPT_FILE=path/to/prompt.md .claude/skills/super-agent/scripts/super-agent
# Inline prompt
.claude/skills/super-agent/scripts/super-agent Use parallel sub-agents to research this repo
# Stdin
cat prompt.md | .claude/skills/super-agent/scripts/super-agent
If agent asks "should I proceed?" instead of implementing: Resume and confirm:
.claude/skills/super-agent/scripts/super-agent --resume <session-id> "Yes, proceed"
The session ID is always printed at the end of output as [session_id: ...].
| Flag / Variable | Description |
|---|---|
--no-chrome | Disable Chrome (enabled by default) |
--no-persist | Don't save session to disk |
--fork, -f | Fork from SESSION_ID context |
--resume, -r | Resume a previous session |
PROMPT_FILE | Read prompt from file |
BYPASS_PERMISSIONS=1 | Full autonomy mode |
SUPER_AGENT_DEBUG=1 | Print cost to stderr |
SUPER_AGENT_CHROME=0 | Disable Chrome (env var form) |
SESSION_ID | Session ID for --fork |
SUPER_AGENT_MODEL | Override model (e.g., claude-sonnet-4-5) |
Transcripts land in ~/.claude/projects/<cwd-slug>/. Use external transcript monitor for realtime visibility.
Wrapper around codex CLI for non-interactive runs (prompt via args/file/stdin) with resume support. Outputs session_id for conversation continuation.
Multi-agent debate orchestrator. Use when you want multiple AI perspectives on a question, plan review, or architectural decision. Runs parallel agent panels with a Chair synthesizing responses.
Wrapper around gemini CLI for non-interactive runs (prompt via args/file/stdin) with resume support. Outputs session_id for conversation continuation.
This skill should be used when the user asks to "implement this plan", "generate execution prompts", "super implement", "create handoff from plan", or wants to transform a large plan into self-contained execution-ready prompt artifacts.
ALWAYS use this skill instead of Read when reading Claude Code transcript files (.jsonl in ~/.claude/projects/, ~/.claude/history.jsonl, or paths containing 'transcript'). Token-efficient format that extracts USER/ASSISTANT exchanges and tool summaries, strips metadata.
Build a phased prompt chain with orchestration file from a monolithic implementation prompt