Claude Code plugins

ClosedLoop.ai distributes its agent and orchestration layer as six Claude Code plugins. Each plugin is a self-contained marketplace entry that installs under ~/.claude/plugins/cache/closedloop-ai/.

The roster

Plugin	Purpose
`bootstrap`	Meta-orchestrator that reads your codebase and generates a project-specific agent suite and critic config.
`code`	The hub plugin. Implements the full ClosedLoop autonomous SDLC — PRD → plan → critics → implementation → tests → visual QA → validation.
`code-review`	Multi-agent partitioned code review with deterministic hygiene checks, risk-based routing, caching, and GitHub PR inline comments.
`judges`	LLM-as-judge framework with 21 judge agents that score plans, code, and PRDs.
`platform`	Claude Code expert guidance, prompt engineering, mermaid diagrams, artifact upload, and skill scaffolding.
`self-learning`	Pattern capture and organizational knowledge sharing. Stores patterns in TOON format with deterministic success rates.

How they fit together

Plugins have implicit cross-dependencies:

code depends on judges (via the eval-cache and run-judges skills).
code depends on self-learning (via learning-quality skill and the 7 Python pipeline scripts called from run-loop.sh).
code-review depends on code (uses code:code-review-worker).
code-review depends on judges (reuses validate_judge_report and JUDGE_REGISTRY).
bootstrap depends on code (uses code:find-plugin-file).

The docs/dependencies.md file in the plugins repo has the full graph.

Slash commands

Commands are namespaced by plugin (/<plugin>:<command>). The full reference is at Commands reference. The most-used commands:

/code:code — start a ClosedLoop coding session against a PRD or plan.
/code:amend-plan — interactively amend a running plan.
/code:cancel-code — stop a running loop.
/code:plan-with-codex — run a Claude-versus-Codex debate to refine a plan.
/code-review:start — run multi-agent partitioned code review.
/self-learning:process-learnings — classify and merge pending learnings.
/self-learning:push-learnings, /pull-learnings — sync org-level patterns.
/bootstrap:start — generate a project-specific agent suite.

The code plugin ships 23 agents (planning, implementation, cross-repo, support). The judges plugin ships 21 judge agents plus a context manager. The bootstrap plugin ships 8 agents for the discovery and generation pipeline.

Agents are Markdown files with YAML frontmatter declaring name, description, model, tools, and skills. Model selection follows a cost/signal convention:

Opus for creative or structural work (plan drafting, high-signal judges).
Sonnet for implementation and validation.
Haiku for lightweight coordination and formatting.

Skills

Skills are deterministic or semi-deterministic playbooks invoked as plugin:skill-name. Examples:

code:plan-validate — Python plan validator (VALID / FORMAT_ISSUES / EMPTY_FILE).
code:critic-cache, code:build-status-cache, code:cross-repo-cache — cache gates.
code:closedloop-env, code:find-plugin-file — environment helpers.
code:iterative-retrieval — 4-phase dispatch/evaluate/refine/loop retrieval protocol.
code:prd-creator — PRD scaffolding with ID conventions (US-###, AC-###.#, Q-###).
judges:run-judges, judges:artifact-type-tailored-context, judges:eval-cache.
self-learning:toon-format, self-learning:learning-quality.
platform:claude-code-expert, platform:context-engineering, platform:mermaid-visualizer, platform:upload-artifact, platform:claude-creator.

Hooks

The code plugin registers lifecycle hooks in plugins/code/hooks/hooks.json:

Event	Script	Purpose
`SessionStart`	`session-start-hook.sh`	Create PID → session mapping.
`SessionEnd`	`session-end-hook.sh`	Clean stale session state.
`SubagentStart`	`subagent-start-hook.sh`	Set up loop state; inject up to 15 relevant org patterns; inject env vars.
`SubagentStop` (order 1)	`loop-stop-hook.sh`	Validate loop agent promise tokens; force retries up to per-agent budget.
`SubagentStop` (order 2)	`subagent-stop-hook.sh`	Enforce learning capture; update `outcomes.log` and `perf.jsonl`.
`PreToolUse`	`pretooluse-hook.sh`	Inject tool-specific patterns (Bash, Write, Edit).

Versioning

Each plugin has its own version in its .claude-plugin/plugin.json manifest. The code@closedloop-ai plugin version is reported in desktop.hello.pluginVersion. Override it for tests with CL_PLUGIN_VERSION (semver, max 64 characters).

Every plugin change must update CHANGELOG.md (enforced by the repo's pre-push hook when git config core.hooksPath .githooks is active).