claude-brewcode

Claude Code plugin suite -- four plugins for development, documentation, text utility, and visual workflows.

A regular Claude Code session hands a big task to one agent and loses sight of it. Brewcode splits work into bounded units, gives every spawn a six-field brief, and re-states the delegation rule on every prompt. Four plugins. 28 skills. 8 agents. 6 lifecycle hooks.

Full Documentation

Approaches -- every hook, agent and injection the suite installs, and who installs what.

Security: Code is scanned on every commit. No personal data is collected or transmitted.

🚀 Install in 30 seconds

Paste this prompt into any Claude Code session -- Claude will run the full install for you:

Execute these commands in this Claude Code session, one by one, show full output for each, do not skip any, do not summarize:

1. claude plugin marketplace add https://github.com/kochetkov-ma/claude-brewcode
2. claude plugin install brewcode@claude-brewcode
3. claude plugin install brewdoc@claude-brewcode
4. claude plugin install brewtools@claude-brewcode
5. claude plugin install brewui@claude-brewcode

After all commands succeed, run `/reload-plugins`. If `/reload-plugins` is unavailable, tell me to type `exit` and run `claude` again. Run the commands now.

Already installed? Update with this prompt:

Execute these commands in this Claude Code session, one by one, show full output for each, do not skip any, do not summarize:

1. claude plugin marketplace update claude-brewcode
2. claude plugin update brewcode@claude-brewcode
3. claude plugin update brewdoc@claude-brewcode
4. claude plugin update brewtools@claude-brewcode
5. claude plugin update brewui@claude-brewcode

After all commands succeed, run `/reload-plugins`. If `/reload-plugins` is unavailable, tell me to type `exit` and run `claude` again. Run the commands now.

After install/update, run /reload-plugins (or exit + claude to restart). For manual install steps see Installation below.

Plugin Suite

Installation

Marketplace (recommended)

Permanent installation through the Claude Code plugin system. Plugins load automatically with every session.

1. Add marketplace -- registers the plugin source:

claude plugin marketplace add https://github.com/kochetkov-ma/claude-brewcode

1. Install plugins -- copies plugins to local cache:

claude plugin install brewcode@claude-brewcode
claude plugin install brewdoc@claude-brewcode
claude plugin install brewtools@claude-brewcode
claude plugin install brewui@claude-brewcode

1. Reload plugins -- run /reload-plugins in Claude Code, or exit + claude to restart.

Already installed? Update

Use /brewtools:plugin-update inside Claude Code for the easiest path -- it runs all update commands and reloads automatically.

Or run manually:

claude plugin marketplace update claude-brewcode
claude plugin update brewcode@claude-brewcode
claude plugin update brewdoc@claude-brewcode
claude plugin update brewtools@claude-brewcode
claude plugin update brewui@claude-brewcode

After updating, run /reload-plugins (preferred) or exit + claude to restart.

Local directory (for development)

claude --plugin-dir ./brewcode --plugin-dir ./brewdoc --plugin-dir ./brewtools --plugin-dir ./brewui

Requirements

macOS only. Linux and Windows support is planned for future releases.

Quick Start

brewcode -- infinite task execution

/brewtools:task-board-setup  # 1. Deploy the task board + a project-tailored /task-spec skill
/task-spec "Implement JWT authorization"  # 2. Research codebase + write the task spec and design
/brewcode:superreview-setup  # 3. Generate a project-tailored deep-review skill
/brewcode:setup-status       # anytime: what is installed, stale or missing in this project

Skills orchestrate, agents execute. Each spawn is a bounded unit with a six-field brief; the forced-eval hook re-states the manager role and the split rule on every prompt, and a role-recall/compact-recall pair re-anchors the same role plus the plan and task graph after a compaction, so work stays observable across compaction cycles.

brewdoc -- documentation tools

/brewdoc:docsync-setup                # Install doc-staleness tracking, then sync stale docs
/brewdoc:my-claude                    # Generate Claude Code installation docs
/brewdoc:memory-sync-setup            # Generate a project-tailored /memory-sync skill into the repo
/brewdoc:md-to-pdf ./docs/report.md   # Convert markdown to PDF
/brewdoc:publish "Hello world"        # Publish to brewpage.app -- returns public URL

brewtools -- text utilities

/brewtools:text-optimize CLAUDE.md         # Token-efficient optimization (52 rules)
/brewtools:text-human 3be67487             # Remove AI artifacts from a commit
/brewtools:secrets-scan                    # Scan for leaked credentials
/brewtools:plugin-update                   # Install or update the plugin suite
/brewtools:context-slim measure            # Weigh the permanent context surface, read-only

brewui -- visual tools

Placeholder plugin, currently empty. No commands yet -- coming soon.

How It Works

  /task-spec "..." --> parallel research agents + user Q&A --> per-task spec + design docs
        │
        v
  project agents from .claude/agents/ --> bounded units, fanned out in ONE message
        │
        v
  /brewcode:superreview-setup --> project-tailored deep-review skill

  every prompt: forced-eval (UserPromptSubmit) injects 3 lines --
    [ROLE]   scan agents, project .claude/agents/ first; domain expert exists -> delegate
    [SPLIT]  one agent for an hour = drift you cannot observe -> split and fan out
    [BRANCH] no branch chosen -> main; no explicit branch/PR -> take over the whole workspace

  after a compaction: role-recall + compact-recall (SessionStart, matcher compact) re-inject
    the same 3 lines plus plan/intent and task graph -- forced-eval never fires there, since
    an auto-compaction has no prompt to hook into

Delegation contract

Every skill that spawns subagents carries a ## Delegation section; every shipped agent carries a ## Scope guard. Both encode the same rule: one subagent = ONE bounded unit -- one deliverable, ~5 files, ~10 steps. Anything bigger is split into N tasks spawned in a single message.

Every spawn prompt carries six fields:

Skills Reference

The -setup suffix marks a skill you run once to install a mechanism -- afterwards you use what it produced (a generated skill, a hook, an MCP server), not the skill itself. Recurring tools you invoke every day keep bare names. Every -setup skill shares one mode vocabulary: status | install | upgrade | enable | disable | uninstall | purge, and no argument means status when installed, install when not. The one exception is /brewcode:semble-setup, which always defaults to status so a bare invocation never triggers a machine-level package install. All eleven setups implement all seven verbs, via one of two mechanisms: a live config flag re-checked each invocation (semble, agent-deadline, agent-return, agent-router, manager, docsync), or entry-file parking, where the filename discovery keys on is renamed <name>.disabled with the body byte-identical (teams, superreview, task-board, think-short, memory-sync).

All 28 skills are user-invoked only. Every one carries user-invocable: true and disable-model-invocation: true in its frontmatter: the model never sees their descriptions and never auto-activates one. You type /plugin:skill, or nothing runs. This is a deliberate trade about context cost -- 28 model-visible descriptions would be a permanent tax on every request -- and these skills do not want auto-activation anyway: eleven of them write real files into your repo after asking you real questions, and the rest are tools you point at a scope you choose.

Brewcode (9 skills)

Brewdoc (5 skills)

Brewtools (14 skills)

Brewui (0 skills)

No skills yet -- placeholder for future UI/visual/creative tools.

Portable skills (standalone)

Self-contained SKILL.md folders that ship outside the four plugins -- drop them into any compatible agent runtime. Both publish content to brewpage.app (text, markdown, files, multi-file sites) and return a public URL.

Agents (8 total)

Scope guard: every shipped agent carries a ## Scope guard -- a task larger than one bounded unit (one deliverable, ~5 files, ~10 steps) makes the agent stop before starting and return a split proposal instead of grinding for an hour. ssh-admin and deploy-admin additionally split per host, repo, and environment.

Documentation

Development

CLAUDE_DEBUG=1 claude --plugin-dir ./brewcode   # Debug mode
bash .claude/scripts/update-plugin.sh           # Update all plugins
bash .claude/scripts/bump-version.sh X.Y.Z      # Bump version everywhere

License

MIT -- see LICENSE