super-plan-mode

 ███████╗██╗   ██╗██████╗ ███████╗██████╗
 ██╔════╝██║   ██║██╔══██╗██╔════╝██╔══██╗
 ███████╗██║   ██║██████╔╝█████╗  ██████╔╝
 ╚════██║██║   ██║██╔═══╝ ██╔══╝  ██╔══██╗
 ███████║╚██████╔╝██║     ███████╗██║  ██║
 ╚══════╝ ╚═════╝ ╚═╝     ╚══════╝╚═╝  ╚═╝

 ██████╗ ██╗      █████╗ ███╗   ██╗    ███╗   ███╗ ██████╗ ██████╗ ███████╗
 ██╔══██╗██║     ██╔══██╗████╗  ██║    ████╗ ████║██╔═══██╗██╔══██╗██╔════╝
 ██████╔╝██║     ███████║██╔██╗ ██║    ██╔████╔██║██║   ██║██║  ██║█████╗
 ██╔═══╝ ██║     ██╔══██║██║╚██╗██║    ██║╚██╔╝██║██║   ██║██║  ██║██╔══╝
 ██║     ███████╗██║  ██║██║ ╚████║    ██║ ╚═╝ ██║╚██████╔╝██████╔╝███████╗
 ╚═╝     ╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝   ╚═╝     ╚═╝ ╚═════╝ ╚═════╝ ╚══════╝

Plan first. Build with confidence.

Built natively for OpenClaw and Claude Code · Works with Cursor, Windsurf, Codex CLI, Aider, and more.

---

What It Is

Super Plan Mode is a universal agent skill that generates a rich implementation plan and halts until you explicitly approve it before writing a single line of code.

Built natively for OpenClaw and Claude Code following the Agent Skills Open Standard. Compatible with any AI coding agent — Cursor, Windsurf, Codex CLI, Aider, and more.

No passive guardrails. No soft suggestions. A hard stop, a structured plan document, and a numbered menu waiting for your decision.

---

Features

• 🛑 Explicit accept/build gate — numbered menu halts execution; nothing runs until you say so
• 📋 Rich plan artifacts — problem statement, affected files, implementation steps, risk matrix, effort estimate, acceptance criteria, rollback notes, and alternative approaches — all in one saved document
• 🟢🟡🔴 Confidence scores — every implementation step rated by how well-understood it is based on codebase exploration
• 🔍 Parallel codebase exploration — multiple read-only agents research architecture, affected files, and risks simultaneously (supported in OpenClaw, Claude Code, Cursor 2.4+, Windsurf Wave 13+, and Codex CLI)
• ⚠️ Pre-flight checks — warns if uncommitted changes overlap with the plan before you accept
• 🔁 Modify loop — type 3 to revise the plan; see a diff of what changed before re-approving
• 💾 Plan persistence — plans saved to .claude/plans/ for reference, sharing, and resuming
• ▶️ Resume saved plans — --resume <file> loads a past plan and jumps straight to the gate
• 📁 Plan history — --list shows all saved plans for the current project
• 🌐 Multi-agent compatible — works across Claude Code, Codex CLI, Cursor Agent, Windsurf, and Aider
• 🤖 Model + provider flag — --model provider/model lets you choose the implementation model from any provider
• 💨 Dry-run mode — --dry-run generates and saves the plan without offering to execute
• ⚡ Process optimized — parallel agents with narrow, non-overlapping scopes; structured output keeps results focused
• ⚙️ Project config — .super-plan-mode.json sets per-project defaults for save path and checkpoints

---

Quick Start

# Full plan → gate → build workflow
/super-plan-mode add OAuth login with Google and GitHub

# Shorthand alias (same thing)
/spm add a payment webhook handler

# Generate plan only, build later
/spm --dry-run refactor the auth middleware

# Resume a saved plan
/spm --resume .claude/plans/super-plan-mode-1234567890.md

# List all saved plans
/spm --list

# Choose implementation model
/spm --model provider/model add a caching layer

Or trigger passively — just tell the agent:

"Let's plan this before writing any code" "Show me the plan first" "Don't touch anything yet, just plan it out"

---

How It Works

Phase 1 — Understand

Clarifies the request. Asks focused questions if anything is ambiguous. Confirms understanding before moving on.

Phase 2 — Explore (Read-Only)

Launches parallel explore agents (where supported) to map the codebase from three angles:

• Architecture — patterns, conventions, existing similar features
• Affected Files — what changes, what depends on it, blast radius
• Risk / Conflicts — uncommitted changes, fragile files, recent churn

In single-agent environments (Cursor, Codex, Windsurf, Aider), the same research runs inline sequentially. Same result either way.

Phase 3 — Generate Plan

Produces a complete plan document and saves it to .claude/plans/. The plan includes:

Phase 4 — The Gate

Pre-flight checks run first (uncommitted file conflicts, test command detection). Then the plan is presented with this menu:

────────────────────────────────────────────
  Plan ready. What would you like to do?

  1  Accept & Build
  2  Reject / Cancel
  3  Modify plan
  4  Accept Phase 1 only  ← (large plans only)
────────────────────────────────────────────
  Enter number (or type the action):

Nothing happens until you respond. Type 3 to iterate; see what changed in the plan before re-approving. Type 2 to cancel cleanly.

Phase 5 — Implementation

Executes the approved plan step by step. For Large/XL plans, pauses at each phase checkpoint to report progress and confirm continuation. Ends with a summary of everything that changed.

---

Confidence Scores

Every implementation step is rated based on how thoroughly it was understood during exploration:

---

Model Flag

Use --model provider/model to specify which model runs the implementation. Omit it to use whatever your environment has active. Any provider/model value is accepted as passthrough — ensure your environment is configured for the model you want to use.

---

Project Config

Add .super-plan-mode.json to your project root to set defaults:

{
  "planSaveDir": ".claude/plans",
  "autoPhaseCheckpoints": true,
  "preflightChecks": true
}

CLI flags always override config file values.

---

Works With

See skills/plan-mode/references/agent-compat.md for detailed per-a