convez.ai
SkillDesenvolvimento WebIntermediário
ralphcrisostomo

LangGraph ADK Skill

ralphcrisostomo/langgraph-adk-skill

Gere CLIs LangGraph acionais com aprovação arquitetural visual

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill do Claude Code que arquiteta e gera CLIs acionais interativas em JavaScript/TypeScript usando LangGraph e Bun. Automatiza a criação de agentes conversacionais com interface terminal moderna (Ink/React), diagrama ASCII de validação e integração com LangSmith.

Instalação

$npx convez-ai ralphcrisostomo/langgraph-adk-skill

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v0.2.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Scaffolding rápido de agentes conversacionais em TypeScript
  • Criar assistentes CLI com rastreamento em tempo real de nós LangGraph
  • Construir CLIs interativas com UI terminal (menus, spinners, prompts contextuais)

Veredito

Skill sólido para quem trabalha com LangGraph e precisa de scaffold rápido com CLI moderna. Forte em arquitetura visual e geração de código TypeScript, mas exige familiaridade com LangGraph e Bun.

README

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

LangGraph ADK

A Claude Code skill that architects and scaffolds JavaScript/TypeScript LangGraph "ADK" (Agent Development Kit) command-line apps — interactive, action-oriented Bun CLIs.

Invoke it and describe what you want. The skill:

  1. Detects the mode — scaffold a new CLI, or add an action to an existing one.
  2. Architects one action — for agent-type actions it analyzes the scenario, composes a fitting LangGraph design from primitives, and renders an ASCII diagram for you to approve before any code is written.
  3. Verifies the current API via context7 (the LangChain JS API drifts) and generates + registers the action.
  4. Gates completion with a Codex review — after verifying, it emits a /codex:review --base=<short-sha> --wait line and won't call the work done until the review has run.

Generated CLIs use minimist (flags fill parameters) and an Ink UI (React for the terminal — menus, text input, spinners; prompts only for what's missing), with chalk for plain colored output and live LangGraph node/tool tracing. The system prompt is stamped with the current date/time each turn (LLMs have no clock). Multi-turn actions render a Claude Code-style Ink <SessionApp>: the input framed by dividers above a ccstatusline-style status footer (working dir · model · context-usage bar), with a configurable PROMPT_LABEL prompt prefix.

Recipes ship for common scenarios — an AWS CLI + repo shell assistant (one pinned account, human-approved writes) and LangSmith tracing.

Install

Clone this repo straight into your Claude Code skills directory:

git clone https://github.com/ralphcrisostomo/langgraph-adk-skill.git ~/.claude/skills/langgraph-adk

Then invoke it in Claude Code:

/langgraph-adk

…or just say "scaffold me a LangGraph CLI that …" and the skill activates.

What a generated CLI looks like

my-cli/
  package.json            # bun; scripts: start, test, typecheck
  src/
    index.ts              # minimist parse -> action select -> run
    cli.ts                # action runner; resolves flags/prompts via Ink
    config.ts             # env -> typed config (local LLM by default)
    dotenv.ts             # load the project's own .env (for the global bin)
    llm.ts                # model factory
    model-info.ts         # resolves the model's real context window
    trace.ts              # verbose LangGraph streaming (streamAgent)
    session-core.ts       # pure session helpers (status footer, date stamp, history trim)
    params.ts             # Ink chooser + flag/prompt resolver
    ui.tsx                # Ink prompts (menu, text input, confirm, spinner)
    ink/
      SessionApp.tsx      # multi-turn Ink chat UI (framed input + dir/model/context footer)
    actions/
      _types.ts           # Action / ParamDef / Ctx contract
      index.ts            # explicit action registry
      chat.tsx            # demo action; `bun run start` works immediately

Run it: bun run start (interactive) or bun run start --action <name> --flag value.

Install as a global command

The generated package.json exposes a bin named after the project, so you can run the CLI from anywhere:

bun link            # once, from the project dir — registers `<project-name>` globally
cd ~/some/other/repo
<project-name>      # runs here; the footer Dir + file tools operate on this dir

A linked bin runs with your shell's current directory, so actions operate on wherever you invoke it (CWD still overrides this). The project's own .env is loaded automatically — resolved from the source location — so your LLM config works no matter where you run it.

Action types

  • Function action — deterministic logic or a single model call.
  • Agent action — a LangGraph graph (single agent, sequential, parallel, loop-and-critic, coordinator/routing, orchestrator-worker, map-reduce, agent-as-tool, or a custom composition). Executes through streamAgent so every node and tool call is rendered live.

Stack

PackageRole
langchaincreateAgent, tool (LangChain v1.x)
@langchain/langgraphgraph primitives, streaming, checkpointing
@langchain/openaiChatOpenAI (+ custom baseURL for local endpoints)
@langchain/anthropicChatAnthropic
ink · @inkjs/ui · ink-text-input · reactterminal UI (menus, input, spinner, chat session)
minimist · chalkflag parsing · colored output

Default LLM config (env)

Defaults to a local, OpenAI-compatible endpoint — override via env:

LLM_PROVIDER=local            # local | openai | anthropic
LLM_MODEL=google/gemma-4-26B-A4B-it
LLM_BASE_URL=http://localhost:8000/v1
# LLM_API_KEY=                 # optional for local endpoints
# PROMPT_LABEL=you             # session prompt prefix, e.g. "you › "
# CWD=.                        # working dir actions run in (default: launch dir)

Run bun link to install the CLI as a global command (named after the package); it runs from any directory, operates on that directory, and still loads this project's .env.

Set LANGSMITH_TRACING=true + LANGSMITH_API_KEY to send run traces to LangSmith (see the observability recipe).

Repo layout

SKILL.md                              # the skill: architect method + scaffold logic
references/primitives.md              # LangGraph building-block cheat-sheet
references/chat-aws-repo.md           # recipe: AWS CLI + repo bash assistant in chat
references/langsmith-observability.md # recipe: add LangSmith tracing to a project
templates/                            # verbatim project skeleton + action templates
reference/                            # proving-ground CLI (TDD'd, `bun test`) — templates derive from this
docs/superpowers/                     # design spec + implementation plan

templates/src is a verbatim copy of reference/src. The reference/ app is the tested source of truth; if you change one, keep the other in sync.

Develop / verify

cd reference
bun install
bun test          # unit tests
bun run typecheck
bun run scripts/scaffold-smoke.ts   # instantiate templates -> install -> typecheck -> --list

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados