Repo Bootstrap

Repo Bootstrap

Make marketplace plugins load in a repo — including web sessions, where ~/.claude doesn't persist. The durable mechanism is a committed .claude/settings.json that declares the marketplace and enables the plugins. This skill writes it (and optional CI).

Local-CLI-only use doesn't need this: /plugin marketplace add cportka/claude-plugins then /plugin install <name>@portka-tools persists in ~/.claude across all local repos.

Steps

1. Ask which plugins to enable (e.g. video-bug-analyzer), whether to add CI, and whether to apply the Portka standard setup (--portka-standard, below). Default marketplace: portka-tools (cportka/claude-plugins). Unattended/autonomous session (nobody to ask)? Use the deterministic default: --portka-standard --scope project --ci (repo-bootstrap enables itself automatically), plus --identity "Name <email>" when the owner's identity is known (#114 con 5).
2. Run the script:

   ${CLAUDE_PLUGIN_ROOT}/skills/repo-bootstrap/scripts/bootstrap-repo.sh \
     --plugin <name> [--plugin <name> ...] [--ci] [--dir <repo-root>] [--dry-run] [--auto-update] \
     [--portka-standard] [--scope user|project|both] [--print-only]
   

   --dry-run previews without writing (and states the real outcome per file). --auto-update is catalog-only on third-party marketplaces — claude plugin update <name>@<marketplace> is the reliable refresh (see --help). Merges never clobber other keys; invalid existing JSON stops the run rather than risk data loss.
3. Review + commit .claude/settings.json (and any workflow). It only takes effect once committed — web sessions clone the repo fresh and read it at session start.

Resulting file:

{
  "extraKnownMarketplaces": {
    "portka-tools": { "source": { "source": "github", "repo": "cportka/claude-plugins" } }
  },
  "enabledPlugins": { "video-bug-analyzer@portka-tools": true }
}

Portka standard setup (--portka-standard)

Installs the contract: describe a feature or next step and it's understood Claude branches fresh from main, builds + tests fully, opens the PR, merges on CI green, and hands the PR back for the user to delete as confirmation. Concretely (all idempotent, never clobbering; full inventory in --help): the managed workflow CLAUDE.md block (version-stamped; the SessionStart hook flags a stale copy), the committed .claude/commit-identity declaration (applied to git config immediately by this run and each session after — the end of identity setup, including in the bootstrapping session itself), a git/gh permissions allowlist, an enforced SemVer sync with tests/run-tests.sh + CI (native node --test/pytest version-sync tests for JS/Python repos), and the corrected stop-hook-git-check.sh that stops hosted sessions false-flagging GitHub's squash-merge commits — reported at session start and installed only on request with --heal-stop-hook (it lives outside the plugin directory, so nothing writes it unasked; a .stock.bak backup is kept). Run that flag when a session starts flagging merged commits as unpushed/unverified.

The decisions the agent must make: which plugins, --scope (user = ~/.claude, project = committed ./.claude for web sessions + team, both = default), and committing the project files so fresh web sessions pick them up. Existing files are never overwritten (test runner only with --force); --home redirects user-scope writes (mainly for tests).

When the .claude/settings.json write is blocked (--print-only)

In a web session under auto mode, Claude Code's permission classifier can refuse an agent-written .claude/settings.json (it flags adding a marketplace + enabling plugins + a permissions allowlist as untrusted self-modification). That's the very file this skill exists to write. When that happens:

1. Ask the user to approve the write explicitly, rather than silently eating the denial.
2. If still blocked, run with --print-only: it prints the exact .claude/settings.json (and, with --portka-standard, the CLAUDE.md workflow block) to stdout. Have the user create/paste the file by hand — a human-authored write isn't classifier-gated — then commit it. The version/sync scaffold isn't classifier-gated, so re-run without --print-only to write it.