convez.ai
SkillDevOps & InfraestruturaAvançado
yongwoon

YWC Agent Toolkit

yongwoon/ywc-agent-toolkit

Automação completa do workflow de desenvolvimento com skills e agentes

Ver no GitHub
BomPluginScore · de 0 a 100

Coleção de 36+ skills para Claude Code e Codex que automatiza o fluxo completo de desenvolvimento, desde planejamento e especificação até geração de código, review e release. Inclui 12+ agentes customizados e suporta integração com GitHub, testes automatizados e merge de dependências.

Instalação

$npx convez-ai yongwoon/ywc-agent-toolkit

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
8
Forks
1
Acompanhando
0
Saúde
71%
Último commit
há 1 semana
Versão
v1.40.0
Issues abertas
4
Contribuidores
2

Casos de uso

  • Automatizar pipeline de desenvolvimento end-to-end
  • Gerar especificações e código de forma coordenada
  • Gerenciar PRs, reviews e releases no GitHub
  • Executar testes e validações em paralelo

Veredito

Toolkit robusto e bem estruturado para automatizar workflows complexos de desenvolvimento. Recomendado para equipes que buscam integração profunda com Claude Code/Codex e GitHub, mas exige familiaridade com shell scripting e configuração de ferramentas.

README

A tradução para o português deste README ainda está na fila.

ywc-agent-toolkit

A collection of skills for Claude Code and Codex that automates the full development workflow — from planning and spec writing to code generation, review, and release.

한국어 | 中文 | 日本語 | Español

Supported Tools

ToolSkillsCustom AgentsInstall path
Claude Code3612~/.claude/skills/, ~/.claude/agents/
Codex377~/.codex/skills/, ~/.codex/agents/

Prerequisites

Plugin marketplace and Codex plugin installation have no prerequisites — the tool handles everything automatically.

For the bash script fallback, the following must be installed before running install.sh:

ToolRequired forInstall
gitCloning the repositoryPre-installed on most systems
bash ≥ 3.2Running install.shPre-installed on macOS / Linux
jqHook registrationbrew install jq / apt-get install jq

At skill runtime (not required for installation):

ToolUsed byInstall
python3 ≥ 3.9Skill runtime helpers: ywc-parallel-executor, ywc-finish-branch, ywc-merge-dependabot; Claude Code hooks require Python ≥ 3.11Pre-installed on macOS 12.3+; brew install python3
gh CLIPR-based and GitHub release skills/modes: ywc-handle-pr-reviews, ywc-spec-writer --from-pr/--from-prs, ywc-release-pr-list, ywc-create-pr, ywc-finish-branch PR mode, ywc-merge-dependabot, ywc-sequential-executor/ywc-parallel-executor, ywc-gen-testcasebrew install gh / cli.github.com

Installation

Via Claude Code plugin marketplace (recommended)

# Add as a marketplace source (one-time setup)
/plugin marketplace add yongwoon/ywc-agent-toolkit

After running the command, open the Plugin UI (Marketplaces tab) and install ywc-agent-toolkit from there. Skills and Claude Code agents are installed automatically — no cloning or bash required.

Via Codex CLI plugin directory

This repository includes Codex plugin packaging metadata under .codex-plugin/, with plugin-local skills mirrored in .codex-plugin/skills/ and checked by bash scripts/validate.sh. This prepares ywc-agent-toolkit for Codex CLI/App plugin installation, but does not mean it is already listed in an official Codex marketplace.

When a Codex plugin marketplace or source that contains this repository is available:

# Shell
codex

# Inside the interactive Codex session
/plugins

Inside the interactive Codex session, open the plugin directory, search for ywc-agent-toolkit from yongwoon/ywc-agent-toolkit, then choose Install plugin.

Via Codex App Plugins sidebar

In the Codex App, open Plugins from the sidebar, search or browse for ywc-agent-toolkit, confirm the plugin source is yongwoon/ywc-agent-toolkit, then install it from the plugin details view.

Until this repository appears in a Codex plugin source available to you, use the bash fallback below.

Via bash script fallback

YWC_REF=<release-tag-or-reviewed-commit>
git clone --branch "$YWC_REF" --depth 1 https://github.com/yongwoon/ywc-agent-toolkit.git
cd ywc-agent-toolkit
git remote get-url origin
git rev-parse --verify HEAD

# Claude Code
bash scripts/install.sh --cc

# Codex
bash scripts/install.sh --codex

# Both
bash scripts/install.sh --all

Install specific skills only

bash scripts/install.sh --cc ywc-plan ywc-commit ywc-create-pr
bash scripts/install.sh --codex ywc-plan ywc-commit ywc-ui-ux-review

Install only custom agents

# All 12 Claude Code worker/reviewer/specialist agents
bash scripts/install.sh --cc-agents

# All 7 read-only specialist agents
bash scripts/install.sh --codex-agents

# Selected agents
bash scripts/install.sh --cc-agents ywc-backend-coder ywc-qa-engineer
bash scripts/install.sh --codex-agents ywc-security-engineer ywc-architect

# Skills only, leaving agents untouched
bash scripts/install.sh --cc --skip-agents
bash scripts/install.sh --codex --skip-agents

List available items

bash scripts/install.sh --list
bash scripts/install.sh --list --cc
bash scripts/install.sh --list --codex
bash scripts/install.sh --list --cc-agents
bash scripts/install.sh --list --codex-agents

Environment variables

VariableDefaultDescription
CLAUDE_SKILLS_DIR~/.claude/skillsOverride Claude Code install path
CLAUDE_AGENTS_DIR~/.claude/agentsOverride Claude Code agent install path
CODEX_HOME~/.codexOverride Codex home path

Restart Claude Code or Codex after installation for skills to take effect.


Quick Start

Install via the Claude Code plugin marketplace — no cloning or bash required:

/plugin marketplace add yongwoon/ywc-agent-toolkit

Then invoke any skill immediately inside Claude Code:

/ywc-onboard-repo           # understand an unfamiliar codebase in minutes
/ywc-plan                   # turn a rough idea into a plan or spec
/ywc-debug-rootcause        # trace a bug to its root cause
/ywc-impl-review            # review code for spec / security / quality
/ywc-agentic                # run the full pipeline autonomously from a goal

Not sure which skill to use? → What do I want to do?


Skills

Most ywc-* skills are available for both Claude Code and Codex. The Codex versions include Codex-compatible frontmatter and tool guidance.

What do I want to do?

GoalSkills
Turn an idea into a plan or specywc-plan → ywc-spec-writer
Understand an unfamiliar codebaseywc-onboard-repo
Break work into dependency-safe tasksywc-task-generator
Implement tasks end-to-endywc-sequential-executor / ywc-parallel-executor
Run the full pipeline from a goalywc-agentic
Find the root cause of a bugywc-debug-rootcause
Review code quality and securityywc-impl-review, ywc-security-audit
Open a PR and handle review commentsywc-create-pr → ywc-handle-pr-reviews
Generate a QA test sheetywc-gen-testcase
Write release notesywc-release-pr-list + ywc-changelog-release-notes
Author a new ywc-* skillywc-skill-author

Planning & Spec

SkillDescription
ywc-planConvert a rough idea into plan.md (Small) or a Spec document (Medium/Large)
ywc-spec-writerWrite and update spec documents (docs/specification/)
ywc-spec-validateValidate spec quality (Completeness / Consistency / Feasibility)
ywc-tech-researchResearch libraries and compare technical approaches
ywc-ubiquitous-languageCreate and maintain a domain ubiquitous language dictionary
ywc-brainstormShape rough ideas before writing a formal plan or spec
ywc-confidence-gateCheck readiness and risk before starting substantial implementation
ywc-onboard-repoGenerate repository onboarding context for unfamiliar projects

Task & Execution

SkillDescription
ywc-task-generatorDecompose a spec into dependency-safe tasks
ywc-sequential-executorExecute tasks sequentially (Branch → Implement → Commit → PR → Merge)
ywc-parallel-executorExecute tasks in parallel using Git worktree isolation
ywc-code-genGenerate Backend + Frontend + QA code in parallel
ywc-agenticAutonomously orchestrate the ywc-* pipeline from a goal (Plan → Execute → Evaluate → Repeat, max 3 iterations)
ywc-tdd-ritualDrive feature and bugfix work through a red-green-refactor loop
ywc-worktreesCreate, audit, prune, and resolve worktree-based task isolation

Review & Verification

SkillDescription
ywc-impl-reviewVerify implementation across Architecture / Design / Devex / Security / QA axes
ywc-review-learningsAccumulate per-project review preferences in docs/review-learnings.md
ywc-security-auditSecurity audit based on OWASP Top 10
ywc-ui-ux-reviewUI/UX review (IA + Visual design + WCAG 2.2 AA)
ywc-design-renewRenew generic UI surfaces and audit AI-slop design tells
ywc-product-reviewProduct feedback across 5 business dimensions
ywc-gen-testcaseGenerate test sheets from PRs or tasks
ywc-e2e-test-strategyDesign Playwright E2E test strategy
ywc-debug-rootcauseInvestigate bugs, failed tests, and build failures to the root cause
ywc-receive-reviewTriage and apply human or automated review feedback
ywc-refactor-cleanRemove dead code, unused exports, stale files, and unused dependencies
ywc-verify-doneVerify tests, builds, and completion evidence before declaring work done

Git & Release

SkillDescription
ywc-commitStage and commit session work
ywc-create-prCommit and create a Draft PR
ywc-handle-pr-reviewsAutomate PR review responses
ywc-finish-branchFull branch delivery (CI → merge → cleanup)
ywc-merge-dependabotAuto-merge Dependabot PRs
ywc-release-pr-listSummarize PRs included in a release
ywc-changelog-release-notesGenerate CHANGELOG entries and user-facing release notes

Documentation & Other

SkillDescription
ywc-project-scaffoldGenerate directory structure for any language/framework
ywc-project-docsGenerate project documentation in Korean or Japanese
ywc-incident-postmortemWrite a structured postmortem after a production incident
ywc-skill-author(Meta) Rules for authoring new ywc-* skills

Codex-only

SkillDescription
ywc-team-assembleSplit explicitly authorized work across specialist Codex subagents

Custom Agents

Claude Code also ships 12 custom agents for worker, reviewer, and specialist dispatch. They install to ~/.claude/agents/ and are documented in claude-code/agents/README.md.

Seven read-only specialist agents complement the ywc-* skills. They are installed to ~/.codex/agents/ (override with CODEX_HOME) as individual TOML files, and Codex loads one custom agent per file.

AgentPurposeSandbox
ywc-architectArchitectural decision and trade-off advisorread-only
ywc-security-engineerStatic security review and threat-model triageread-only
ywc-root-cause-analystRoot-cause and incident-cause analysisread-only
ywc-performance-engineerPerformance review and profiling recommendationsread-only
ywc-typescript-reviewerTypeScript / JavaScript language-specific reviewread-only
ywc-python-reviewerPython language-specific reviewread-only
ywc-go-reviewerGo language-specific reviewread-only

All Codex agents are read-only; they return a standardized Status: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT, a compact verdict or finding set, and a Next action: when the caller should apply or inspect something. They never edit files. Source TOML lives under codex/agents/.


HTML Output Mode for Review Skills

Nine review and report skills support an opt-in --format html flag that produces a self-contained, browser-ready HTML report instead of Markdown.

Supported skills: ywc-impl-review, ywc-security-audit, ywc-spec-validate, ywc-tech-research, ywc-incident-postmortem, ywc-product-review, ywc-ui-ux-review, ywc-gen-testcase, ywc-design-renew

Why HTML? AI-generated Markdown documents longer than ~100 lines are rarely read end to end — an unread report cannot drive a decision. HTML adds color, severity coding, tabs, and interactive controls (checkboxes, Copy as Markdown), so the human on the other end actually reads and acts on the output.

/ywc-impl-review --spec docs/spec.md --code src/ --format html
/ywc-security-audit --code api/src/ --format html
/ywc-gen-testcase 250 --format html   # interactive testsheet with localStorage sign-off

⚠️ Token cost — HTML output uses 2–4× the output tokens of Markdown and takes longer to generate. The default is markdown; enable HTML only for reports a human will read in a browser.

Details: references/html-output.md.


Recommended Development Pipeline

This spine mirrors how the skills are actually invoked day to day, not the full catalog. One planning pass, a spec gate, task decomposition, then the executor as the workhorse — for each task it delivers end to end via ywc-finish-branch, folding conformance review (--review), PR creation, bot-review handling, and merge in as sub-steps, so those rarely run standalone in the task-driven flow.

flowchart TD
    A["1. ywc-plan\nrough idea → plan.md"] --> B{Size?}
    B -->|Small| D["3. ywc-task-generator\ndecompose into tasks"]
    B -->|"Medium / Large"| C["2. ywc-spec-writer\nwrite / update spec"]
    C --> CV["ywc-spec-validate\ngate spec quality"]
    CV --> D
    D --> E["4. ywc-sequential-executor\nor ywc-parallel-executor\nbranch → impl → verify → PR → merge"]
    E --> F["5. ywc-gen-testcase pr N\nQA test sheet per PR"]
# Step 4 example — run a task range with full delivery:
ywc-sequential-executor 000020-010..000025-010 --review --base-branch <feature>
# common flags: --base-branch · --draft · --local-merge · --review · --per-task-pr
# (ywc-parallel-executor is the worktree-isolated alternative)

Ad-hoc / non-task changes skip the executor and deliver manually: ywc-create-pr opens a draft PR, then ywc-handle-pr-reviews drives bot / human review to green. ywc-handle-pr-reviews is also what you re-run whenever new review comments land on an open PR — task-driven or not.

Also reached for in real work: ywc-ubiquitous-language (domain glossary, before or during spec), and at release time ywc-release-pr-list + ywc-changelog-release-notes.

The remaining skills are situational, not part of every run — ywc-debug-rootcause (a test or build fails and the cause is unclear), ywc-tdd-ritual (strict red-green-refactor), ywc-tech-research (compare approaches before deciding), ywc-impl-review (standalone conformance review outside the executor), ywc-agentic (autonomously orchestrate the whole pipeline from a goal), and others in the Skills table above.


Claude Code Hooks

Automation hooks that run before/after Claude Code tool calls. Hooks are installed to ~/.claude/hooks/ (global) or ./.claude/hooks/ (project-local) and registered in settings.json automatically.

Install hooks

# Install all hooks globally (default)
bash scripts/install.sh --hooks

# Install all hooks into the current project
bash scripts/install.sh --hooks --local

# Install specific hooks only
bash scripts/install.sh --hooks block-dangerous-commands cost-tracker
bash scripts/install.sh --hooks --local session-start

List available hooks

bash scripts/install.sh --list --hooks

Available hooks

HookEventDescription
block-dangerous-commandsPreToolUseBlock dangerous shell commands (critical/high/strict levels)
check-claude-md-freshnessPreToolUseVerify CLAUDE.md is up to date before git push
cost-trackerPostToolUse + StopLog tool call stats and print session summary on exit
notify-permissionNotificationSend a Slack alert when Claude is waiting for permission (CCH_SLA_WEBHOOK required)
permission-requestPermissionRequestAuto-approve safe tools (Read, Write, Edit)
protect-secretsPreToolUseBlock access to .env, SSH keys, and other secret files
session-startSessionStartInject git status, CONTEXT.md, TODOs, and GitHub Issues at session start

Dependencies

DependencyRequiredInstall
jqYes — JSON mergebrew install jq / apt-get install jq
uvYes — run Python hookscurl -LsSf https://astral.sh/uv/install.sh | sh

For per-hook usage details see claude-code/hooks/README.md.


Contributing

Contributions are welcome! Please read CONTRIBUTING.md before submitting a PR.

License

MIT

Ver o original no GitHub

Bom para

Relacionados