aep-onboard

Onboard

Set up your environment for agentic TypeScript development AND get oriented to how AEP thinks. Phase 0 gives you the 5-minute mental model; Phases 1–5 install the plugin, verify tools, and configure recommended plugins. Run once on first setup — returning users running for environment verification can skip Phase 0 and jump straight to Phase 1.

---

Phase 0 — Orient Yourself (5 minutes, first-timers only)

Returning user? If you've run /aep-onboard before and you're just re-verifying your environment, skip to Phase 1.

Before installing tools, get the mental model. AEP is not a "command runner" — it's a workflow that separates thinking (what to build) from doing (building it). Installing the tools without understanding this will leave you staring at a blank terminal wondering which of 16 skills to run first.

The three mental models you need:

1. Control plane vs execution plane. You + AI make high-leverage decisions on the control plane (goals, decomposition, architecture, priorities). Agents receive precise specs and build on the execution plane. They never share code context directly — only structured artifacts like product-context.yaml and signal files. See README.md "The Mental Model".

2. The story map. Your product is organized as a Jeff Patton story map — a grid with activities (columns, user journey left→right), layers (rows, enrichment top→down), waves (parallel batches within a layer), and release lines (what's shippable). Layer 0 is the walking skeleton — the thinnest end-to-end path. See README.md "The Story Map".

3. Two-session model. The main session runs on your integration branch (main in single-branch mode, or develop in two-branch mode — see Phase 5) where you + AI plan (/aep-envision, /aep-map, /aep-dispatch, /aep-design, /aep-wrap, /aep-reflect). The workspace session runs autonomously in an isolated git worktree on a feat/<name> branch where one agent implements a feature (/aep-build). They communicate only through signal files in .dev-workflow/signals/. See skills/product-context/README.md.

v2 split-mode (good to know): Some projects store product context in two files — product/index.yaml (stable intent: opportunity, personas, capabilities, constraints) + product-context.yaml (mutable state: architecture, stories, cost, changelog). All skills auto-detect which mode a project uses. If you see only product-context.yaml, that's v1 single-file mode and it works exactly the same way. See docs/aep-v2-improvement-guideline.md.

Next step: for the full 10-minute first-hour guide — including a table of all 16 skills, four concrete paths (new product / existing project / single feature / hands-free), and a glossary shortlist — read docs/orientation.md. Then come back to Phase 1.

---

Phase 1 — Install the Plugin

Install the AEP skills with the skills CLI at project level, once per agent your project uses. Pin to the latest release and commit the installed files so the version is frozen for your team:

# Claude Code (repeat with `-a codex` for Codex). Newest tag:
# https://github.com/memorysaver/agentic-engineering-patterns/releases/latest
npx skills add memorysaver/agentic-engineering-patterns@<latest-tag> -a claude-code --skill '*' -y

This installs every AEP skill (the aep-* names) plus a skills-lock.json manifest — commit both. For the full pinning + formatter guidance, see Installing Skills.

Optional add-ons — always ask the user

AEP pairs with two project-level skills from memorysaver/skills. Ask the user whether they want each, and install only what they choose (newest tag at https://github.com/memorysaver/skills/releases/latest, once per agent):

• Behavioral guidelines in AGENTS.md? → install project-behavior, then run it to scaffold/extend AGENTS.md.
• A project memory system (committed lessons + recall)? → install project-memory (and memory-forge), run project-memory to bootstrap project-memory/, then add a concise ## Memory & Learning Loop section to AGENTS.md that layers these onto AEP's native lessons loop instead of duplicating it. AEP already captures (/aep-build → .dev-workflow/lessons.md), archives (/aep-wrap → lessons-learned/), and recalls (/aep-launch); the supplement adds: project-memory recall at /aep-dispatch + persisting the archived lesson at /aep-wrap (qmd semantic recall), and memory-forge distilling settled lessons (≥7 days, ≥3 accrued) into skills at /aep-reflect / pre-PR.

npx skills add memorysaver/skills@<latest-tag> -a claude-code \
  --skill project-behavior --skill project-memory --skill memory-forge -y

Note: This installs the AEP skills themselves. Recommended third-party Claude Code plugins are configured at the project level in Phase 4 via .claude/settings.json; browser automation is added only after its local smoke test passes.

---

Phase 2 — Verify Required Tools

Each tool below earns its place in the agentic workflow — git provides version control and worktrees (one isolated working tree per parallel agent), bun runs the TypeScript monorepo, openspec powers spec-driven development, an executor (claude or codex) runs the implementation agents, and gh publishes PRs. tmux is optional: launches are native-first (Claude Code agent teams / background sessions, Codex subagents — see aep-executor); tmux only hosts the pinned legacy mode and the generic-host fallback.

Run this check:

# Required: at least one executor (claude OR codex)
command -v claude >/dev/null 2>&1 || command -v codex >/dev/null 2>&1 \
  && echo "executor:      OK" || echo "executor:      MISSING (install claude or codex)"

# Required: everything else
for cmd in bun git gh openspec; do
  printf "%-15s" "$cmd:"
  which $cmd >/dev/null 2>&1 && echo "OK ($(which $cmd))" || echo "MISSING"
done

# Optional (legacy/pinned-tmux mode only): tmux
printf "%-15s" "tmux:"
which tmux >/dev/null 2>&1 && echo "OK ($(which tmux))" || echo "MISSING (optional — only the legacy launch mode needs it)"

Install any missing tools:

Tool	Purpose	Install
git	Version control + worktrees	xcode-select --install (macOS)
bun	Package manager & runtime	curl -fsSL https://bun.sh/install | bash
claude	Executor: Claude Code CLI	npm install -g @anthropic-ai/claude-code
codex	Executor: OpenAI Codex CLI	npm install -g @openai/codex (alt to claude)
gh	GitHub CLI for PRs	brew install gh
openspec	Spec-driven development (Node >= 20.19)	npm install -g @fission-ai/openspec@latest
tmux	Terminal multiplexer (optional — legacy mode)	brew install tmux

All required tools (executor + bun/git/gh/openspec) must show OK before proceeding. You need at least one executor (claude or codex) — not both. tmux may be MISSING; that's fine — launches are native-first.

Native-first launches: the executor abstraction picks the host's native mode automatically — Claude Code agent teams (claude-team) or background sessions (claude-bg); Codex native subagents (codex-subagent) or exec workers (codex-exec). All include live monitoring and steering without tmux. See aep-executor and the "Enable the native launch modes" step in Phase 5.

Note on parallelism: Each parallel feature agent runs in its own git worktree at .feature-workspaces/<name>/ on its own feat/<name> branch. Worktrees share the underlying .git/objects (no history duplication) but each adds one full working-tree copy on disk — budget accordingly when running many agents in parallel.

---

Phase 3 — Verify Optional Tools

for cmd in cmux agent-browser portless; do
  printf "%-15s" "$cmd:"
  which $cmd >/dev/null 2>&1 && echo "OK ($(which $cmd))" || echo "MISSING (optional)"
done

Tool	Purpose	Install
cmux	Clickable tab multiplexer for watching legacy-mode tmux sessions — optional; only used when tmux is pinned	bun add -g cmux
agent-browser	Browser automation testing	Claude Code plugin: agent-browser@agent-browser
portless	Port management (.localhost)	bun add -g portless

cmux is a convenience, not a requirement. It only adds clickable tabs for watching legacy-mode tmux sessions (when aep.executor-backend tmux is pinned). Without it, pinned workspaces still run in tmux with the full monitor + mid-flight-feedback loop — attach with tmux attach -t <name>. Skills auto-detect cmux and never abort when it's absent. See aep-executor.

These are optional — the workflow works without them but is enhanced by them. On macOS, do not enable agent-browser until a one-command smoke test can launch a page without crashing Chrome:

agent-browser navigate about:blank

If macOS shows a Google Chrome crash report with _RegisterApplication, TransformProcessType, or abort() called, leave agent-browser disabled and use non-browser checks (curl, unit tests, screenshots from the user, or the host agent's browser tool) until the local Chrome/agent-browser combination is healthy.

---

Phase 4 — Configure Project Plugins

Configure recommended plugins at the project level. These plugins are not optional cosmetics — superpowers provides the planning/TDD skills that /aep-design assumes exist, mgrep powers deeper search, frontend-design is assumed by visual calibration work, code-review is used by /aep-build, and the hooks enforce the concurrency protocol that keeps parallel workspace agents from corrupting product-context.yaml.

What to write

Read .claude/settings.json if it exists. Merge the following keys into it (or create the file if missing):

{
  "extraKnownMarketplaces": {
    "claude-plugins-official": {
      "source": { "source": "github", "repo": "anthropics/claude-plugins-official" }
    },
    "superpowers-marketplace": {
      "source": { "source": "github", "repo": "obra/superpowers-marketplace" }
    },
    "Mixedbread-Grep": {
      "source": { "source": "github", "repo": "mixedbread-ai/mgrep" }
    }
  },
  "enabledPlugins": {
    "superpowers@superpowers-marketplace": true,
    "frontend-design@claude-plugins-official": true,
    "code-review@claude-plugins-official": true,
    "mgrep@Mixedbread-Grep": true,
    "skill-creator@claude-plugins-official": true
  },
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path // \"\"' | { read -r f; case \"$f\" in *product-context.yaml) if [[ \"$PWD\" == */.feature-workspaces/* ]]; then echo '{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"permissionDecision\":\"deny\",\"permissionDecisionReason\":\"CONCURRENCY PROTOCOL: Workspace sessions must not write to product-context.yaml. Write to .dev-workflow/signals/status.json instead. Only the main session (via /aep-wrap, /aep-dispatch, /aep-reflect) updates the YAML.\"}}'; fi ;; esac; }",
            "statusMessage": "Checking concurrency protocol..."
          }
        ]
      },
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.command // \"\"' | { read -r cmd; if [[ \"$PWD\" == */.feature-workspaces/* ]] && echo \"$cmd\" | grep -qE 'git\\s+(add|commit).*product-context\\.yaml'; then echo '{\"hookSpecificOutput\":{\"hookEventName\":\"PreToolUse\",\"permissionDecision\":\"deny\",\"permissionDecisionReason\":\"CONCURRENCY PROTOCOL: Workspace sessions must not commit product-context.yaml. Write to .dev-workflow/signals/status.json instead.\"}}'; fi; }",
            "statusMessage": "Checking concurrency protocol..."
          }
        ]
      }
    ]
  }
}

Concurrency protocol hooks: The hooks section enforces the rule that only the main session writes to product-context.yaml. When a workspace agent attempts to edit, write, or commit product-context.yaml, the hook blocks the action and explains how to use signals instead. This is defense-in-depth — the skill instructions also direct agents to use signals, but the hook catches any model drift.

Optional browser automation

Only add agent-browser after the Phase 3 smoke test succeeds. It launches a local Chrome process, and some macOS/Chrome combinations crash during application registration before the test can run.

{
  "extraKnownMarketplaces": {
    "agent-browser": {
      "source": { "source": "github", "repo": "vercel-labs/agent-browser" }
    }
  },
  "enabledPlugins": {
    "agent-browser@agent-browser": true
  }
}

Plugin reference

Plugin	Marketplace	Purpose
superpowers	superpowers-marketplace	Planning, debugging, TDD, code review workflows
agent-browser	agent-browser	Optional browser automation for testing
frontend-design	claude-plugins-official	High-quality UI generation
code-review	claude-plugins-official	PR code review
mgrep	Mixedbread-Grep	Semantic search (local + web)
skill-creator	claude-plugins-official	Create and test new skills

Merging rules

• If .claude/settings.json already has these keys, merge new entries — do not overwrite other keys
• Preserve any existing settings (permissions, env, etc.)
• If .claude/settings.json already has a hooks.PreToolUse array, append these hook entries — do not replace existing hooks
• If the file doesn't exist, create it with all three keys (extraKnownMarketplaces, enabledPlugins, hooks)

---

Phase 5 — Verify Environment

echo "=== Core Tools ==="
command -v claude >/dev/null 2>&1 || command -v codex >/dev/null 2>&1 \
  && echo "executor:      OK" || echo "executor:      MISSING (claude or codex)"
for cmd in bun git gh openspec; do
  printf "%-15s" "$cmd:"
  which $cmd >/dev/null 2>&1 && echo "OK" || echo "MISSING"
done
echo ""
echo "=== Optional Tools ==="
for cmd in tmux cmux agent-browser portless; do
  printf "%-15s" "$cmd:"
  which $cmd >/dev/null 2>&1 && echo "OK" || echo "MISSING (optional)"
done
echo ""
echo "=== Git Repo ==="
[ -d .git ] && echo "git repo: OK" || echo "Not a git repo — run: git init"
git worktree list 2>/dev/null | head -5

If all core tools show OK, the environment is ready.

Integration branch (single- vs two-branch mode)

AEP integrates all feature work into one integration branch ($BASE across the skills). The standard cases are auto-detected — you do not configure anything. Report which mode this repo is in:

# Auto-detect (same logic every skill uses): develop → two-branch; otherwise single-branch
if git show-ref --verify --quiet refs/heads/develop \
   || git show-ref --verify --quiet refs/remotes/origin/develop; then
  echo "Integration branch: develop  (two-branch mode — main is promote-only production)"
else
  echo "Integration branch: main  (single-branch mode)"
fi

• single-branch mode (no develop): AEP integrates into main. The default; matches a brand-new repo — don't create develop just for AEP.
• two-branch mode (develop exists): AEP integrates into develop (staging); production main is promote-only and AEP never touches it. Promotion develop → main is your CI/CD or PR step — exactly like deployment, which AEP leaves to you.
• No pinning, no reconfiguration. Because the standard cases are auto-detected, a project grows from single- to two-branch mode simply by creating develop — every skill picks it up automatically on the next run. Do not run git config aep.integration-branch main; pinning the default would suppress that upgrade.
• Non-standard name only: if your integration branch is neither main nor develop (e.g. staging, integration), set the override once: git config aep.integration-branch <name>. It lives in .git/config, shared across all worktrees. See git-ref → "Integration Branch".

Enable the native launch modes (recommended)

/aep-launch and /aep-autopilot pick the launch mode automatically (see aep-executor): native first, tmux only when pinned. Two one-time setup steps unlock the best native modes:

Claude Code — agent teams (claude-team mode). Agent teams is still experimental (flag-gated, Claude Code ≥ 2.1.32). Enable it per-project via the env block of .claude/settings.json (a settings entry, not a hook — AEP's no-hooks stance is unaffected). Ask the user before writing:

{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

Without the flag, launches fall back to claude-bg (native background sessions, GA) — fully functional, with pull-based steering.

Codex — custom agent roles (codex-subagent mode). Commit the two AEP role files into the project's .codex/agents/ so both the desktop app and the CLI discover them (the TOML contents are in aep-executor/references/codex-native.md):

mkdir -p .codex/agents
# write aep-builder.toml and aep-evaluator.toml from the codex-native.md templates
git add .codex/agents && git commit -m "chore: add AEP codex agent roles"

Prefer the old tmux+cmux workflow? Pin it: git config aep.executor-backend tmux.

---

Next Steps — Pick Your Path

Your next move depends on your situation. Pick the path that matches what you're trying to do. Full context for each path (including why each step is in the order it's in) is in docs/orientation.md section 4, "The Four Paths".

Path A — New product from scratch

You have an idea and a fresh repo.

/aep-envision  →  /aep-map  →  /aep-validate  →  /aep-scaffold  →  /aep-autopilot

/aep-envision validates the opportunity and extracts the activity backbone. /aep-map decomposes it into a system map + story graph + agent topology. /aep-validate runs gen/eval checks. /aep-scaffold creates the monorepo + OpenSpec. /aep-autopilot (optional) takes over hands-free — or drive it manually with /aep-dispatch → /aep-design → /aep-launch → /aep-build → /aep-wrap.

Path B — Onboarding an existing project

You have a codebase and want to add AEP workflows to it.

/aep-scaffold  →  /aep-dispatch  →  /aep-design  →  /aep-launch  →  /aep-build  →  /aep-wrap

/aep-scaffold adds agentic infrastructure (OpenSpec, workspace hooks, E2E skeleton) to existing code. Then start a feature cycle with /aep-dispatch. Use /aep-envision later if you want to retrofit a product context.

Path C — Single feature, no product context

You just want to ship one feature with AEP workflows.

/aep-design  →  /aep-launch  →  /aep-build  →  /aep-wrap

/aep-design produces an OpenSpec change on the integration branch ($BASE). /aep-launch spawns an isolated git worktree on a feat/<name> branch and boots the agent. /aep-build implements, tests, reviews, and merges. /aep-wrap archives and removes the worktree.

Path D — Hands-free autonomous mode

You have a validated product context and want to go grab coffee.

/aep-autopilot

One command. Pauses only for design escalations or layer gate failures. Deep dive: docs/workflow/autonomous-loop.md.

Still unsure which path? See the decision tree in docs/skills-quick-reference.md.

---

Guardrails

• Run from the project root — tools and plugins are verified relative to the current environment
• Re-run anytime — safe to re-run to verify environment is still complete. Returning users can skip Phase 0 (orientation) and jump to Phase 1.
• Checks only — this skill verifies and installs tools, it does not scaffold projects or modify code

---

Learn More

Pointers for going deeper. None of these are required reading — check what's relevant when you need it.

Mental models & concepts

• docs/orientation.md — the canonical first-hour guide (mental models + 16 skills + four paths)
• README.md "Why This Exists" — the full argument for spec-precision-over-execution-speed
• docs/glossary.md — precise definitions for every AEP term (ubiquitous language)

Skills decision guide

• docs/skills-quick-reference.md — cheat sheet + decision tree + common sequences

Autonomous mode

• docs/workflow/autonomous-loop.md — how /aep-autopilot orchestrates dispatch → launch → monitor → wrap

v2 upgrades

• docs/aep-v2-improvement-guideline.md — split-mode, capability maps, readiness scoring, outcome contracts, technical specs, grouped changes

Git + worktree conventions

• skills/agentic-development-workflow/git-ref/SKILL.md — AEP git + worktree reference (worktree lifecycle, branch naming, commit-per-task pattern, recovery), accessed on-demand via /aep-git-ref