marketplace-dev — build, audit, and maintain Claude Code plugins

marketplace-dev gives plugin authors the scaffolding, audit, and self-maintenance infrastructure the dev-team plugin developed internally — as reusable, installable tooling. It targets three workflows: creating a new plugin from scratch, improving an existing plugin's architecture and quality, and establishing a new marketplace with correct structure from the start.

It encodes the conventions in docs/marketplace-builder-plugin-playbook.md — directory layout, agent/skill frontmatter contracts, the markdown-vs-script decision framework, and the eval-fixture pattern — as scaffolding skills, a structural review agent, and a single shared knowledge file.

marketplace-dev has no hard runtime dependency on dev-team.

Slash commands

Command	Role	What it does
/scaffold-plugin <name>	implementation	Create a new plugin dir with the audit-clean skeleton (plugin.json, CLAUDE.md, install.sh, settings.json, agents/ skills/ hooks/ knowledge/)
/scaffold-marketplace <owner>	implementation	Create a marketplace root — catalog, release-please wiring, ≥1 plugin slot
/init-plugin-eval <name>	implementation	Scaffold evals/<name>/{fixtures,expected}/ + a grading-contract README (empty corpus grades as a clean no-op)
/agent-type-advisor <prose \| file>	worker	Recommend markdown vs script for a use-case (forward-looking) or audit an existing agent/skill file (retrospective), with rule citations
/plugin-audit [dir] [--fix]	orchestrator	Generalized structural compliance for any plugin — agent type, frontmatter, eval coverage, body budgets (architecture only)
/agent-create	worker	Create an agent file following the official schema and token budgets
/agent-skill-authoring	worker	Conventions, anti-patterns, and meta-patterns for authoring agents and skills
/agent-add	implementation	Create a new review or team agent (delegates to agent-create)
/agent-remove	implementation	Remove an agent and clean up its registry/doc references
/add-plugin <name@marketplace>	implementation	Install a plugin and register it in a project's settings.json

Agent

• plugin-best-practices-review — read-only, JSON output, structural findings. Checks agent type appropriateness (markdown vs script), frontmatter compliance, eval-coverage presence, and body line-count budgets. It does not evaluate detection-logic quality — that belongs to the plugin's own agent-eval.

Knowledge

• knowledge/agent-type-decision-rules.md — the markdown-vs-script decision matrix (rules R1–R10). The single source of truth shared by /agent-type-advisor (forward-looking) and plugin-best-practices-review (retrospective). Cite rule IDs; never restate.

Conventions enforced

• Shipping hygiene. Only shipped files live under plugins/<name>/. Eval fixtures and tests live at the repo root (evals/<name>/, tests/) — never inside a plugin dir.
• Independent versioning. Each plugin carries its own semver in its plugin.json; release-please keeps plugin.json, the release tag, and the catalog entry in lock-step. Do not hand-edit versions.
• Portability. All shell is #!/usr/bin/env bash and bash 3.2-safe across macOS, Linux, and Git Bash on Windows; every install.sh carries the Git-Bash-on-Windows guard.
• Audit-clean bar. Scaffolded plugins and migrated skills must pass /plugin-audit with zero findings; plugin-best-practices-review produces zero findings against marketplace-dev itself (dogfood).

Eval fixtures

Live at evals/marketplace-dev/ in the repo root (not shipped). Run the structural check with python3 scripts/eval_grade.py --check-corpus --expected-dir evals/marketplace-dev/expected --fixtures-dir evals/marketplace-dev/fixtures.