New Project

---
name: new-project
description: Graduate an idea from your lab into its own full-rig project workspace under ~/Projects/<slug> (stamped from the lab template — recall engine, hooks, git drift-gate, vendored identity, seeded memory, genome skeletons). Use when the user says "graduate X", "spin up a project for X", or "/new-project".
---

# Graduate a project from the lab

Stamps a self-contained workspace from your lab's `template/` via `new-project.sh`. Do NOT hand-create the structure — run the script; it's the maintained source.

## Step 0 — Gather the essentials (ask one at a time if missing)
- **slug** — lowercase, hyphenated (e.g. `my-app`). Becomes the dir + recall `source`.
- **display name** — human name (e.g. "My App").
- **one-line purpose** — what it is / why it exists (goes into the registry + the project's `AGENTS.md` constitution + `Log/STATUS.md`).
- **full genome?** — default is INTENT + SPEC only; pass `--full` to also scaffold USER-STORIES / INCEPTION / ARCHITECTURE-OPEN-QUESTIONS (only when the project clearly warrants them).
- **harness?** — default is **Claude only**. If the user also wants **Codex** on this project, note it (we pass `--harness claude,codex` in Step 1). Surfaced like `--full`: don't interrogate on a routine graduation — only raise it when Codex actually comes up.

## Step 1 — Confirm, then stamp
Show the user the exact command and confirm (it creates a new repo + a memory namespace — an external action). Run it from your lab's root:

```bash
bash new-project.sh <slug> --name "<Display Name>" --purpose "<one-liner>"   # add --full if agreed
# to also stand the project up for Codex, add:  --harness claude,codex
# to also copy your lab's own memories into the project, add:  --seed-from-lab
```

The script: clean-copies the template; vendors the lab's identity file as `identity/IDENTITY.md`; substitutes placeholders; sets `source=<slug>`; creates `.venv` + installs `sqlite-vec`; seeds the curated starter habits (`memory-seed/*.md`, next to the script) into the project's Claude Code memory namespace — never overwriting a file already there (`--seed-from-lab` also copies your lab's own memories, same no-clobber rule; `--no-seed` skips seeding); `git init` + installs the drift-gate + initial commit; reindexes recall; registers the project in your lab's `Projects-REGISTRY.md`. With `--harness claude,codex` it also dogfoods the new project's own `bootstrap.sh` (after substitution, before the commit) to emit the **git-ignored** Codex layer — recall + identity + ceremonies + apply_patch file-protection + the `lab-reader`/`lab-reviewer` helper agents — the same layer the Lab itself gets.

## Step 2 — Verify the stamp (don't assume)
```bash
DEST=~/Projects/<slug>
ls "$DEST"/{CLAUDE.md,recall.config.json,Source/INTENT.md,Source/SPEC.md}
"$DEST"/.venv/bin/python "$DEST"/tests/test_recall.py   # expect: all tests pass
( cd "$DEST" && bash scripts/recall.sh "how should I handle uncertainty" )   # a seeded memory is recallable
( cd "$DEST" && git log --oneline -1 && ls -la .git/hooks/pre-commit )       # commit + gate symlink
grep -n "<slug>" Projects-REGISTRY.md                                         # registered
```

**If you stamped `--harness claude,codex`,** also confirm the Codex layer wired (it's generated + git-ignored, so it won't show in `git status`):
```bash
ls "$DEST"/{.codex/hooks.json,.codex/agents/lab-reader.toml,.codex/agents/lab-reviewer.toml,.lab/harnesses}   # Codex hooks + helper agents wired, harness state recorded
```
Then tell the user about the **one-time `/hooks` trust**: Codex ignores a project's hooks until you trust them — in an interactive Codex session *inside the project*, approve them in the `/hooks` review; recall + the `apply_patch` file-protection hook fire from then on, not before. Don't claim the Codex hooks are working until that trust step is done.

## Step 3 — Fill the genome (the point of graduating)
Open `Source/INTENT.md` + `Source/SPEC.md` and fill them WITH the user (interview, one question at a time) — or at least seed them from what they've already said about the idea. A graduated project with empty genome docs isn't done. Then update `Log/STATUS.md` for session 1 and let the user drive the first build slice.

## Notes
- **Adding/dropping a harness later:** each project ships its own `bootstrap.sh`. Run *inside the project*, `bash bootstrap.sh --harness claude,codex` re-wires it (or `--harness claude` drops Codex) — the same command also stands a project up after cloning it to another machine. This is per-project and run in that project's own tree; `/new-project` and the Lab never reach back to retro-upgrade an already-graduated project.
- Identity is **vendored** (a copy) and stays one — nothing re-pulls it. To propagate a later identity change, copy the file across by hand.
- **Engine updates, though, are covered:** each project ships its own `update.sh`. Run *inside the project*, `bash update.sh` refreshes machinery (scripts, tests, skills, hooks) from the latest published release and leaves the personal layer alone; `bash update.sh --check` is the read-only look. Opt-in, never automatic — a project changes only when someone runs it there.
- The project is fully walled: its own repo, its own memory namespace. It never reads/writes the lab or any other workspace.
- For a structural dry-run with NO side effects (no venv/seed/git/registry): `new-project.sh <slug> --root /tmp --no-venv --no-seed --no-git --no-reindex`.