new-project

• Outside the plugin. The path in step 2 still reads ${CLAUDE_PLUGIN_ROOT} literally →
• The JDK. java -version below 21, or no java → stop: the hooks and /init-project both
• Inside a project. The current directory holds pom.xml or build.gradle → stop. The

Follow § 2 of ${CLAUDE_SKILL_DIR}/../arch-adopt/SKILL.md — that section owns how the source is found, which ref is the default, and how its commit is resolved. Three bindings, none of which changes what it does:

• There is no project yet, so its third case is the one that applies: the plugin's copy is
• SRC for that section's own download is ${TMPDIR:-/tmp}/nerviz-new-project, set
• --ref from this skill's arguments passes through unchanged. --source does not exist here:

What comes out: the ref, and the commit it names — or none, which § 2 says how to state.

SRC="$PWD/nerviz-<ref, every / replaced by ->"

The project is generated at ../<artifactId> of that tree — this directory — exactly as from a clone. Then, in order:

1. $SRC exists and its own .claude/schemas/extensions.json names a source.fetch_marker
2. $SRC exists otherwise → stop, and name it. Never delete, overwrite or merge into it.
3. $SRC does not exist → step 4.

Every other nerviz-* directory here is an earlier fetch. List them in the report as safe to delete; never delete one.

Run § 3 of ${CLAUDE_SKILL_DIR}/../arch-adopt/SKILL.md with the SRC of step 3, with one binding: its tarball is $TGZ, not $SRC.tgz. SRC here is a directory the person owns, and a ~50 MB archive beside it would stay there, since this skill never runs rm. Replace both "$SRC.tgz" in that section with "$TGZ":

TGZ="${TMPDIR:-/tmp}/nerviz-new-project-fetch.tgz"

Then confirm the tree is a source and not an error page that happened to extract:

ls "$SRC/.claude/skills/init-project/SKILL.md" "$SRC/.claude/hooks/ArchHook.jar"

Either missing → stop, and say the ref holds no init-project.

Read source.fetch_marker from $SRC/.claude/schemas/extensions.json — the fetched copy, never a path from memory — and write that file under $SRC with Write:

{"ref": "<ref>", "commit": "<commit from step 2>"}

Leave commit out when step 2 resolved none. export reads this file when it writes the project's .claude/, and stamps that ref and commit; without it the stamp reads working-tree and unknown, since the tree is no repository. A fetched extensions.json with no source.fetch_marker is a ref older than the marker: write nothing, and say in the report that the stamp will read working-tree.

Print the report below and stop. Never start claude yourself, and never run /init-project in this session (top of this file).

✅ <fetched | reused> — ref <ref> (<short commit | commit unknown>)

Source ....... <base_url host>
Tree ......... <$SRC>
Projects ..... generated here, at <this directory>/<artifactId>
Stamp ........ <ref and commit recorded | working-tree — this ref predates the fetch marker>
Older trees .. <none | the nerviz-* directories, safe to delete>

Next, in a terminal:
  cd "<$SRC>" && claude
Accept the folder's trust prompt: its hooks are what enforce the architecture during the
generation. Then, inside that session:
  /init-project
To approve the write outside that tree once, first: mkdir "<this directory>/<name>", then
/add-dir ../<name>. When it ends, develop in a new session inside the project:
  cd "<this directory>/<artifactId>" && claude

Form 2 (disable-model-invocation: true): it downloads a source tree into the person's directory, so it runs only when they ask for it by name — interview axis 2 of claude-code-architect-designer, a command typed by someone who has the plugin and no clone. The closest rejected form was generating inline from the fetched tree, in this session: nothing of that tree loads here, so parity with /init-project would rest on the model translating every path, and a project that differs from what a clone generates was the one outcome the interview ruled out. Runs on sonnet with effort: low: it follows arch-adopt's fetch, writes one small file and prints.