source-command-plan

Use this skill when the user asks to run the migrated source command plan.

Command Template

/plan — Structured Planning Pipeline

Usage

/plan [--fast | --hard | --parallel | --two] <description>

Auto-detect mode if no flag given:

• Fast — single-file change, familiar pattern, ≤ 2 components
• Hard — multi-file, unfamiliar domain, security-sensitive, or ≥ 3 phases
• Parallel — Hard + per-phase file ownership map
• Two — Hard + 2 approaches for user to select

---

Step 0 — Scope Challenge

Before spawning any agents, challenge scope inline:

# Scope Challenge (Step 0):
#   Exists? → [does this feature already exist in the codebase?]
#   Minimum? → [smallest impl that satisfies requirements]
#   Complexity? → [Fast | Hard] — reasons: multi-file? unfamiliar? security?
#
# Mode: [Fast | Hard | Parallel | Two]

If scope is too large: suggest splitting and wait for user confirmation.

---

Step 1 — Research (Hard / Parallel / Two only)

Spawn 2 plan-researcher agents in parallel:

• Instance A — primary approach and best practices
• Instance B — alternative approach and tradeoffs

// spawning 2 plan-researcher agents in parallel
//
// Researcher A: [approach] → [verdict]
// Researcher B: [approach] → [verdict]

---

Step 2 — Plan Creation

Spawn the planner agent with feature description + mode + research reports.

Agent writes:

plans/YYMMDD-{slug}/
  plan.md
  phase-01-{name}.md
  phase-02-{name}.md
  ...

Two mode: planner produces plan-a.md and plan-b.md — show a summary of both, then wait for user to choose before continuing to Step 3.

// spawning planner agent
//
// Created:
//   plans/{date}-{slug}/plan.md
//   plans/{date}-{slug}/phase-01-{name}.md
//   plans/{date}-{slug}/phase-02-{name}.md

// [Two mode only] → show plan-a and plan-b summaries
// [Review Gate] → "Which approach do you prefer — A or B?" — waiting...

---

Step 3 — Red-Team Review (Hard / Parallel / Two only)

Spawn the plan-reviewer agent with paths to all plan files.

Adjudicate each finding:

• ACCEPTED → edit the relevant plan file immediately
• NOTED → append to Risks section of plan.md
• REJECTED → document reason

// spawning plan-reviewer agent
//
// Security:    "{finding}" → ACCEPTED → phase-02 updated
// Assumption:  "{finding}" → NOTED    → added to risks
// Failure:     "{finding}" → ACCEPTED → plan.md updated
// Verdict: WARN — 0 CRITICAL, 1 HIGH resolved

If plan-reviewer returns BLOCK: revise the flagged phase and re-run before proceeding.

---

Step 4 — Validation + Cook

Ask 3–5 targeted questions about the plan's riskiest points. Wait for user answers.

Then hydrate tasks via TodoWrite:

// T1: {Phase 1} (no blockers)
// T2: {Phase 2} (blocked by T1)
// T3: {Phase 3} (blocked by T2)

Output the exact cook command:

Ready to cook:
/cook /abs/path/plans/{date}-{slug}/plan.md

---

Agents

Agent	Step	Modes
plan-researcher	1 — research (×2 parallel)	Hard, Parallel, Two
planner	2 — creates plan files	All
plan-reviewer	3 — red-team review	Hard, Parallel, Two

---

Integration

• /cook <plan-file> — implement phase by phase
• /code-review — review implementation after cooking
• /fix --quick — fix build errors during cook