My LLM Agent Workflow

Agent-first workflows for Claude Code and OpenCode. This fork of shinpr/claude-code-workflows adds QA-focused workflows, security guardrails, and cross-platform plugin support.

Release Note

llm-agent-workflow is the renamed major release baseline for this repository.

• Marketplace install target is now @llm-agent-workflow.
• Core fork-qualified plugins (dev, qa, env-guard) moved to 1.0.0-jcc.1.
• Independently versioned plugins moved to 1.0.0.

Tag Policy

To keep releases predictable and avoid accidental major-version drift:

• The canonical repository tag baseline is v1.0.0 for this rename release.
• Do not create v2.x tags unless there is an intentional breaking-change release.
• For fork-qualified plugin releases, follow v<upstream>-jcc.<n> tags (for example, v1.0.0-jcc.1).
• For independently versioned plugins, follow plain SemVer tags (for example, v1.0.0).

Quick Start

# clone marketplace repo
git clone https://github.com/jcchikikomori/llm-agent-workflow <install-path>

# inside Claude Code
/plugin marketplace add <install-path>
/plugin install dev@llm-agent-workflow
/plugin install qa@llm-agent-workflow
/plugin install env-guard@llm-agent-workflow
/reload-plugins

# run a workflow
/recipe-implement "Add user authentication with JWT"

Platform support: macOS, Linux, WSL, native Windows.

Plugin Catalog

Plugin	Category	Purpose
skills-md	skills	Language/framework coding rules (Ruby, Python, React, Node.js, Docker, more)
dev	agentic-coding	Main development workflows for design, planning, implementation, and review
qa	product-quality	Acceptance test generation, integration/E2E testing, browser QA
env-guard	behavior-control	Blocks accidental .env and secret exposure
claude-attribution	governance	Anti-AI-slop + attribution governance on external posts and commits
markdown-format	quality-enforcement	Auto-fixes markdown lint issues after writes
commit-guard	behavior-control	Hands commit-message git commands to the user to run, then watches for the result
gh-issue-to-pr	workflow-orchestration	Drives one GitHub issue from investigation to PR lifecycle
memory-guard	behavior-control	Watches sensitive docs and routes changes to memory + remove/stash policy
token-saver	behavior-control	Enforces token-efficient prompting and session hygiene
opencode-migrate	workflow-orchestration	Migrates Claude Code setup into OpenCode
mempalace-docker	behavior-control	Runs MemPalace through Docker as one shared, GPU-aware HTTP hub for every session
ruby-lsp	quality-enforcement	ruby-lsp + RuboCop diagnostics after every Ruby edit, advisory Reek smells, Docker-first
markdown-lsp	quality-enforcement	rumdl LSP diagnostics after every Markdown edit, skill-derived fallback config, Docker-first
resource-guard	behavior-control	Gates heavy work under memory pressure, freezes and resumes background sessions' work and containers, caps each session's docker MCP/LSP containers, WSL-aware
eta	behavior-control	Live ETA band above the prompt for long tool calls, background shells, agents and workflows, from past-run medians or Claude's ~3m hint
wandavision	quality-enforcement	Deterministic image analysis via mcp-vision
metronome	behavior-control	Keeps workflows procedural and step-driven
discover	product-quality	Turns ideas into evidence-backed PRDs
caveman	behavior-control	Token-light response style plugin

External Plugins

metronome, discover, and caveman are pulled from their own repos (URL-sourced entries in .claude-plugin/marketplace.json) and maintained by their original authors. See Contributing for submission rules.

How It Works

Development Workflow

Stage	Primary Agents	Output
Intake	requirement-analyzer	Scope and workflow path (small/medium/large)
Discovery	codebase-analyzer, prd-creator	Context and requirements for implementation
Design	technical-designer, ui-spec-designer	Architecture and testable design docs
Planning	work-planner, task-decomposer	Ordered, commit-ready tasks
Execution	task-executor, task-executor-frontend	Working code per task
Quality	quality-fixer, code-verifier, code-reviewer	Test/lint/type fixes and doc-to-code verification

Diagnosis Workflow

Stage	Primary Agents	Output
Investigate	investigator	Execution-path map and failure candidates
Validate	verifier	Confirmed root causes with evidence
Solve	solver	Tradeoff-based fix options and action steps

Workflow Recipes

All entry points use /recipe-*.

Development Recipes (dev)

Recipe	Purpose
/recipe-implement	End-to-end feature development
/recipe-task	Focused change or bug fix
/recipe-design	Produce design documentation
/recipe-plan	Convert design docs to task plan
/recipe-build	Execute an existing task plan
/recipe-review	Validate code against design docs
/recipe-diagnose	Root-cause investigation and solution path
/recipe-reverse-engineer	Generate PRD/design docs from existing code
/recipe-pr-review	Review an external PR with local codebase context
/recipe-generate-claude-md	Generate a CLAUDE.md from scratch for a project that lacks one
/recipe-update-doc	Update existing docs to stay consistent with code changes

Fullstack and frontend work isn't a separate set of recipes — /recipe-implement routes to a fullstack path automatically when a task spans both backend and frontend, dispatching technical-designer-frontend, task-executor-frontend, and quality-fixer-frontend alongside their backend counterparts.

QA Recipes (qa)

Recipe	Purpose
/recipe-add-integration-tests	Add integration/E2E tests to existing features
/recipe-backend-integration-test	Integration tests scoped to backend endpoints
/recipe-web-qa	Browser-level QA on running web apps

Specialized Agents

plugin-dev (27 agents)

Capability	Agents
Intake and scoping	requirement-analyzer, scope-discoverer
Codebase and requirements	codebase-analyzer, prd-creator, rule-advisor
System and UI design	technical-designer, technical-designer-frontend, ui-spec-designer, design-sync
Planning and decomposition	work-planner, task-decomposer
Implementation	task-executor, task-executor-frontend
Quality and review	quality-fixer, quality-fixer-frontend, code-verifier, code-reviewer, pr-reviewer, security-reviewer, document-reviewer
Diagnostics	investigator, verifier, solver
Knowledge and project memory	context-keeper, context-scouter, claude-md-generator, pr-creator

plugin-qa (4 agents)

Capability	Agents
Test generation	acceptance-test-generator, api-endpoint-tester
Test quality review	integration-test-reviewer
Browser QA	web-qa-reviewer

OpenCode

OpenCode-compatible files are included for supported plugins (plugins/*.ts, commands/*.md, and selected custom agents). Use:

./setup-opencode.sh --global

Use --project <path> for project-local install and --dry-run to preview.

Detailed Plugin Docs

Deep setup and internals live in each plugin README:

• plugin-wandavision/README.md
• plugin-mempalace-docker/README.md
• plugin-memory-guard/README.md
• plugin-opencode-migrate/README.md
• plugin-token-saver/README.md
• plugin-attribution/README.md
• plugin-commit-guard/README.md
• plugin-markdown-format/README.md
• plugin-gh-issue-to-pr/README.md
• plugin-ruby-lsp/README.md
• plugin-markdown-lsp/README.md
• plugin-resource-guard/README.md
• plugin-eta/README.md

Recommended Permissions

Pre-approve common inspection commands in ~/.claude/settings.json to avoid constant permission prompts:

{
  "permissions": {
    "allow": [
      "Bash(git remote:*)",
      "Bash(git status:*)",
      "Bash(git diff:*)",
      "Bash(git log:*)",
      "Bash(gh repo:*)",
      "Bash(gh pr:*)"
    ]
  }
}

For inline PR review posting, include GitHub MCP write permissions too.

Contributing

For marketplace/plugin contribution rules, see CONTRIBUTING.md.

License

MIT. See LICENSE.

Built and maintained by @jcchikikomori.