mj-smart-run

Smart Run — Plan with Opus, Execute with Sonnet

How this skill works

When invoked, you orchestrate two distinct phases:

1. PLAN phase — spawn a single Opus agent to think deeply and produce a structured plan
2. EXEC phase — spawn one or more Sonnet agents to implement the plan

---

cmux 환경 지원

cmux 터미널($CMUX_SOCKET_PATH 설정됨)에서 실행 시 진행 상황을 사이드바에 표시한다:

# Phase 시작 시 호출
[ -n "$CMUX_SOCKET_PATH" ] && cmux set-status "smart-run" "running" --icon "gear"
[ -n "$CMUX_SOCKET_PATH" ] && cmux set-progress 0.0 --label "Smart Run 시작..."

각 Phase 전환 시:

• Phase 1 시작: cmux set-progress 0.2 --label "PLAN: Opus 분석 중..."
• Phase 2 시작: cmux set-progress 0.5 --label "EXEC: Sonnet 실행 중..."
• Phase 3 완료: cmux set-progress 1.0 --label "완료" + cmux notify --title "Smart Run 완료" --body "[태스크 요약]"

---

Phase 1: PLAN (Opus)

Spawn ONE agent with model: "opus" to analyze the task and return a structured plan.

Prompt the Opus agent with:

You are a senior architect. The user wants: <task>

Produce a PLAN with these sections:
## Goal
One-sentence summary.

## Steps
Numbered list. Each step must specify:
- What to do
- Which skill or tool to use (if applicable)
- Input/output dependencies with other steps
- Whether it can run in PARALLEL with other steps (mark with [PARALLEL])

## Risks
Key risks or blockers to watch for.

## Definition of Done
How to verify the work is complete.

Be thorough but concise. This plan will be handed directly to execution agents.

Wait for the Opus agent to return the plan before proceeding.

---

Phase 2: EXEC (Sonnet)

Read the plan from Phase 1. For each step:

• Independent steps (marked [PARALLEL]): spawn Sonnet agents simultaneously in a single message
• Sequential steps: spawn Sonnet agents one at a time, passing the previous result as context

Spawn each execution agent with model: "sonnet".

Prompt each Sonnet execution agent with:

You are an expert implementer. Execute this specific step from a larger plan:

## Your Step
<step description>

## Full Plan Context
<full plan from Opus>

## Prior Step Results (if any)
<results>

Execute completely. Return: what you did, files changed, and any output the next step needs.

---

Phase 3: REPORT

After all execution agents complete, summarize:

• What was planned (Opus)
• What was executed (Sonnet agents)
• Any steps that failed or need follow-up

---

Invocation

When the user runs /smart-run <task> or /플랜실행 <task>:

1. Confirm you understood the task (one line)
2. Announce: "[PLAN] Thinking with Opus..."
3. Run Phase 1 (Opus agent)
4. Show the plan to the user, ask for approval or proceed automatically at L2+
5. Announce: "[EXEC] Executing with Sonnet..."
6. Run Phase 2 (Sonnet agents)
7. Report results