Run, monitor, recover, and babysit ML evaluation experiments end-to-end. Covers the full lifecycle: intent alignment, dry-run validation, auto-launch, checkpoint recovery, real-time/cron-based self-healing monitoring, and post-completion validation. TRIGGER when: user wants to run/start/launch/rerun experiments or evaluations, monitor running experiments or check experiment logs, recover from failures like 429 quota errors or process death, set up monitoring for long-running tasks, or do a dry-run before committing to a full run. Also trigger for 'babysit', 'experiment-runner', and when user provides monitor CLI args like 'monitor experiments/X.json --log Y --expected N'. DO NOT trigger for: analyzing/comparing existing results (tables, charts, LaTeX), writing scripts to parse result files, code review or code modification, dataset creation or preprocessing, GPU/system monitoring, training job submission (use amlt-run-job), or autonomous development tasks.
Bootstrap a new project for Claude Code: writes CLAUDE.md, research-notes.md, pyproject.toml/equivalent, and .gitignore from project-type signals (ML research / web app / CLI / library / data pipeline). Dry-run preview before writing. TRIGGER when: bootstrap project, init-project, 配置新项目, new project setup. DO NOT trigger when: editing an existing CLAUDE.md (use claude-md-improver).
Compact the current conversation into a handoff document for another agent to pick up. Use when the user wants to hand off the current session to a fresh agent, save context before compact, or asks for '交接'/'移交'/'handoff'/'下一个 agent 接手'. Complements save-session (which writes to docs/history/), while handoff writes to OS tmp dir for cross-session pickup.
Use Codex to review code changes AND optionally execute fixes. Two modes — (1) review-only (默认，输出 findings 不动代码); (2) review-and-fix (autonomous, 找出 BLOCKER/MAJOR 后直接派 codex exec 修，最多 N 轮，每轮 Claude 重审). TRIGGER when 用户要 review 代码 / audit a diff / inspect uncommitted work / check branch vs base / 用 codex 修 bug / "review 完顺便修" / "review and fix" / "review 并改" / "review 后自动修"; user mentions /codex-review or invokes Codex for code work. DO NOT trigger for: 单纯写新代码 / feature implementation without prior review intent (let user pick the right entry point), broad architecture discussion without concrete diff.
Use when user invokes /ultra-plan to run an end-to-end loop — interview (Claude clarifies + writes spec) → execute (Codex 默认执行: Single/Todo/Teams 三种粒度，对应 1 次/多次串行/多次并行 `codex exec`) → evaluate (Claude review + AC check) → fix-loop (找到问题派 Codex 修，最多 2 轮) → report. Default Autonomous mode: Phase 1 talks spec clear then AI runs without interrupting user until final report. `/ultra-plan --collaborative` switches to step-by-step user confirmation. **Role split (since 2026-05-10)**: Claude = planner + judge; Codex = executor + fixer. 角色对称。
Simplifies and refines code for clarity, consistency, and maintainability while preserving all functionality. Focuses on recently modified code unless instructed otherwise.
OpenClaw expert — use for ANY OpenClaw question or problem including: config errors, gateway crashes, gateway not starting, slash commands, channel routing, multi-agent setup, heartbeat, cron jobs, channel setup, Discord/Telegram/Slack/iMessage integration, pairing, auth, nodes, version upgrades, skills, sandboxing, model providers, webhooks, hooks, "openclaw not working", "gateway crashed", "command not found", session issues, memory, browser tool, CLI usage, installation, permissions, tool policy, remote access, and any other OpenClaw topic.
Use when user invokes /interview to clarify task requirements before planning or execution. Helps AI understand user's true intent through structured questioning, producing a requirements summary for subsequent planning.