Nested Repos

Nested repos — one git repo living inside another

You have a self-contained sub-project (a plugin, a shared lib, a third-party skill, a repo with its own GitHub remote) and it must ALSO live inside this repo. There are four ways to wire that, and they are NOT interchangeable — the sharp differences are worktree auto-presence, who owns the history, and how you develop/update it. Two classic failures this skill exists to kill: reaching for git submodule by reflex (→ an empty dir in every fresh worktree), and assuming .worktreeinclude will carry a nested repo into a worktree (it will NOT — see leaf 3). Pick from the tree, don't default.

Decision tree — first match wins

1. Third-party code you'll freeze + slim into THIS trunk (drop heavy media, curate a snapshot)? → vendor-and-commit. This leaf already has a skill: use vendoring-skills for the full procedure (placement, slimming, provenance, the lint B3 allowlist). Don't re-derive it here.
2. Must be present in every worktree with ZERO setup, and you mainly CONSUME it from upstream (rarely edit it in place)? → subtree.
3. You DEVELOP it in place as its own repo (push to its own remote), want it in EVERY worktree, and the parent should track only a reviewed source ref? → gitignore + materialize-hook (register it in worktree-repos.json).
4. Parent must pin an EXACT version and you bump deliberately, and a one-time init step per worktree is acceptable? → submodule.

Discriminators at a glance:

The four options

submodule

Parent records a gitlink: .gitmodules (URL) + a pinned commit SHA. The sub-repo stays fully independent (own remote, own history).

• Update: cd <path> && git pull, then in the parent git add <path> && git commit to record the new SHA. Deliberate, never automatic — that pin is the feature (reproducibility).
• Worktree gotcha: git worktree add does NOT check out submodules (Claude Code uses default git worktree logic per the live docs), so the path is EMPTY in a fresh worktree until git submodule update --init --recursive. Storage is shared via .git/modules, so the init is fast — but it IS a per-worktree step.
• Pick when: exact-version pinning matters and the init step is acceptable.

subtree

The sub-repo's files become REAL tracked files in the parent — no nested .git, no gitlink.

• Update: git subtree pull --prefix=<path> <url> <branch> --squash (and git subtree push … to send parent-side edits upstream).
• Worktree: present natively everywhere, zero setup.
• Pick when: you consume from upstream and want frictionless worktree presence; you accept subtree-command sync instead of plain git push.

gitignore + materialize-hook ← the way to ride a NESTED REPO into worktrees

The sub-repo stays a LIVE nested repo on disk (its own .git), .gitignored so it's out of the parent's history.

• .worktreeinclude does NOT work for this. It copies gitignored files, and git does not recurse into a nested-repo boundary, so a nested-repo dir enumerates as nothing to copy. Verified 2026-06-20 (prediction 55b1735b miss — brand-foundry never rode in, twice). Single gitignored files ride fine; a nested repo never does. Do not reach for .worktreeinclude here.
• Instead, REGISTER it. At the repo root, add a {path, remote, ref} entry to worktree-repos.json, where ref is the reviewed full 40-character commit SHA. hooks/materialize_worktree_repos.py (wired to SessionStart
  • PostToolUse[EnterWorktree]) clones any missing entry into the worktree — from the local primary checkout if present (fast, offline), else the remote, checks out that exact detached commit, verifies it, then points origin at the remote. No-op in the primary checkout; never clobbers an existing dir; fails open. A mutable local experiment may explicitly set development: true, but that mode is not reproducible or distribution-safe and must not ship unnoticed.
• Update: develop/git pull in place exactly as a standalone repo.
• Worktree: cloned in automatically on EnterWorktree (verified end-to-end 2026-06-20 — a real EnterWorktree auto-fired the hook and cloned brand-foundry in) and on claude --worktree launch (SessionStart wiring; same engine, not separately live-fired). Subagent isolation:worktree worktrees fire neither trigger — minor edge; clone manually if a subagent needs it.
• Pick when: you develop it in place as its own repo, want it in EVERY worktree, and the parent should track a reviewed source revision without absorbing the sub-repo's files/history. This is what brand-foundry uses.

vendor-and-commit

Slim a clone and commit a snapshot into the trunk at <path>. → skill vendoring-skills owns this procedure; go there. Native in every worktree; updates = re-vendor (never hand-edit, or you lose upstream-trackability).

Worked example — brand-foundry (this harness, 2026-06-20)

skills/brand-foundry/ is its own repo (github.com/GhostlyGawd/brand-foundry), developed in place; no separate clone exists on disk. Goal: develop in place + ride into every worktree, without the trunk owning its history. → tree node 3: gitignore skills/brand-foundry/ (keeps it out of the trunk) + register its remote and reviewed commit in worktree-repos.json; the materialize hook clones and verifies that commit in each worktree. The FIRST attempt used .worktreeinclude and failed — it can't carry a nested-repo dir (prediction 55b1735b miss). Submodule was rejected (empty in fresh worktrees

• trunk-coupling); subtree was rejected (would break in-place own-repo dev).

Packaging a MULTI-skill distributable — use the plugin format

The decision tree picks the git-NESTING mechanism; this picks the PACKAGING when the sub-project is a bundle of several skills (plus its own commands/agents) meant to be distributed and invoked as a unit. Package it as a Claude Code plugin, not as skills hand-nested under one skills/<name>/ dir:

• Give it a .claude-plugin/plugin.json manifest (name is the only field the validator REQUIRES; it only WARNS on missing version/description/author, so add those for a release-quality plugin). Put its skills at <plugin>/skills/<skill>/SKILL.md, and use ${CLAUDE_PLUGIN_ROOT} for any path a component references inside itself. Validate with claude plugin validate <path>.
• A plugin's skills are invoked namespaced as <plugin>:<skill> (e.g. prospector:validate). The namespacing is mandatory and isolating — a plugin skill is NOT merged into the bare top-level skill namespace. (Verified against the Claude Code plugins reference, 2026-06-23.)
• To ALSO develop the plugin in place as its own repo and ride it into every worktree, combine this with decision-tree node 3 (gitignore + materialize-hook via worktree-repos.json) — packaging and nesting are orthogonal choices.

(session 16494681, 2026-06-20: placing the 5-skill prospector bundle needed a dedicated claude-code-guide agent to recover the plugin spec from scratch, because no skill covered multi-skill packaging.)

Rules

• Name the discriminator out loud before picking: worktree auto-presence?, who owns history?, develop-in-place or consume? — those decide it.
• Don't reach for submodule just because "repo inside a repo" sounds like one; it's the only option that does NOT auto-populate a worktree.
• Never use .worktreeinclude to carry a NESTED REPO — it's for gitignored files only. Use the registry + materialize hook.
• This skill is the chooser. Mechanics live elsewhere: vendoring-skills (vendor-and-commit), worktree (.worktreeinclude for gitignored files), and worktree-repos.json + hooks/materialize_worktree_repos.py (nested repos).