Roadmap: repo provisioning + first field review

[DanMat]

Captures two inputs: creating the remote repo as part of scaffolding, and the first unsolicited review from an agent that used Packkit to build a real internal full-stack tool.

The connecting thread

Packkit assumes a greenfield empty directory. The actual workflow is "the repo already exists" — either because you created it, or because you cloned it.

--here aborts when the target has any entry at all, and dirIsEmptyOrMissing counts dotfiles — so a bare .git/ from a fresh clone is enough to stop it. That's why the reviewer had to back up, move, and restore files, and it's the same wall you hit creating an empty repo and cloning it first. Repo provisioning and a merge mode are two halves of one gap.

Added to Next up

• Create the repo, not just the folder — --github [owner/name] via the gh CLI (so Packkit never handles a token), --git-remote <url> as the universal escape hatch. Opt-in, private by default. Side benefit: the repo URL currently gets typed by hand and feeds package.json links + README badges, so deriving it makes existing output correct.
• Scaffold into a non-empty directory — merge mode that writes what's absent and never clobbers.

Added from the review

• Full-stack monorepo preset — the highest-value item. monorepo hardcodes two library packages with no apps/; node-service and react-app already exist, so it's the composition that's missing.
• packkit.json provenance — Packkit leaves zero footprint today, so a project can't say what it started from. Prerequisite for any future packkit upgrade.
• One-line MCP install — largely closed by the registry publish already; what's left is a deep-link install button.
• Preset discovery for agents — the review asked for a tool that exists under a different name (packkit_preview); the misremembering is itself the finding.
• Post-scaffold checklist — extend what the CLI already prints.

Deliberately not acted on

Recorded with reasoning so they don't get re-raised: the "Node 22 vs 24 friction" is the engine preflight working as designed (it refuses to scaffold a project whose tooling can't run, and prints the exact fix), and the "overwrite risk" in --here is inverted — it aborts rather than overwrites.

🤖 Generated with Claude Code