convez.ai
AgenteProdutividadeIntermediário
intincrab

Sidequest – Aprenda enquanto seu agent trabalha

intincrab/sidequest

Reclaime o tempo morto: aprenda enquanto Claude Code trabalha

Ver no GitHub
InicialPluginScore · de 0 a 100

Painel terminal que transforma o tempo de espera de agentes de código (Claude Code) em micro-lições de um modelo LLM econômico. Mantém memória persistente por tópico, continuando exatamente onde parou, e alerta quando o agent termina.

Instalação

$npx convez-ai intincrab/sidequest

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

Casos de uso

  • Preencher downtime de execução de agentes de código com aprendizado contínuo
  • Explorar tópicos diversos (história, ciência, idiomas) em sessões fragmentadas
  • Manter contexto educacional persistente através de múltiplas interrupções

Veredito

Solução criativa que aproveita tempo ocioso real de workflows com agents, via integração com Claude Code. Requer setup manual nas hooks, suporta múltiplos providers baratos (GLM, Haiku, GPT mini), e persiste aprendizado por tópico — bem executado para públicos técnicos.

README

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

sidequest

Learn something while your coding agent is busy.

When you fire off a long task in Claude Code (or Codex), there's dead time while it works — sometimes seconds, sometimes minutes. sidequest is a small terminal panel that reclaims that time to teach you a topic of your choice — the fall of Rome, how black holes work, Spanish basics, the bond market — in bite-sized lessons from a cheap LLM, with its own persistent memory so each wait picks up exactly where you left off. When the agent finishes, it nudges you back.

┌─ Claude Code ─────────────┐ ┌─ sidequest ────────────────────┐
│  > refactor the auth flow │ │  📚 how black holes work · L28 │
│    …working…              │ │  agent ● busy — learn while…   │
│                           │ │                                │
│                           │ │  Black holes might be the      │
│                           │ │  universe's way of creating    │
│                           │ │  life by regulating galaxy …   │
│                           │ │  › /next  /deeper  /topics     │
└──────────┬────────────────┘ └───────────────▲────────────────┘
           │  SessionStart / tool hooks        │ localhost event
           └───────────────────────────────────┘

It's the dead time, reclaimed. The lessons are unrelated to your code — that's the point.


How it works

  1. The panel runs a tiny localhost listener (127.0.0.1:4317 by default).
  2. Claude Code hooks (installed into settings.json) fire a fire-and-forget script on every UserPromptSubmit / tool use / Stop, telling the panel when the agent is busy vs. idle. The script never blocks the agent (300 ms timeout, always exits 0).
  3. While the agent is busy, the panel auto-serves the next bite-sized lesson on your topic.
  4. Lessons come from a pluggable cheap model — Zhipu GLM, Anthropic Haiku, or an OpenAI mini model.
  5. Each topic's conversation is saved under ~/.sidequest/topics/, so you can stop and resume any time — days later it continues mid-thought.

Requirements

  • Node.js 18.18+
  • An API key for one provider (GLM is free-tier friendly — see below)
  • Claude Code (the agent side)
  • Windows Terminal — only needed for the one-command side-by-side pair layout (everything else works in any terminal)

Install

git clone https://github.com/<you>/sidequest.git
cd sidequest
npm install
npm link            # optional: makes `sidequest` available globally

Then wire it into Claude Code (merges hooks into your global ~/.claude/settings.json):

npm run install:claude -- --global

Configure

Pick a provider and give it a key. Copy .env.example to .env:

# .env  (gitignored — never commit this)
SIDEQUEST_PROVIDER=glm
GLM_API_KEY=your-key-here
ProviderSet SIDEQUEST_PROVIDERKey env varDefault modelNotes
Zhipu GLMglmGLM_API_KEYglm-4.5-flashOpenAI-compatible; free tier
AnthropicanthropicANTHROPIC_API_KEYclaude-haiku-4-5—
OpenAIopenaiOPENAI_API_KEYgpt-4o-mini—

Optional settings live in ~/.sidequest/config.json:

{
  "provider": "glm",
  "model": "glm-4.5-flash",
  "defaultTopic": "how black holes work", // auto-loads on launch
  "lessonStyle": "concise and vivid, ~130 words, one memorable detail",
  "autoAdvanceOnBusy": true,               // serve the next lesson when the agent is busy
  "autoLaunch": false                      // auto-open the panel on Claude Code start (Windows Terminal)
}

Usage

One command, side by side (recommended)

From any project folder:

sidequest pair               # Claude Code (this folder) + sidequest, split in one window
sidequest pair C:\some\proj  # …or target a specific folder

This opens a single Windows Terminal window split into two panes: your agent on the left, sidequest docked on the right.

Or run them separately (any terminal)

# pane A
npm start          # the sidequest panel
# pane B
claude             # your agent, as usual

TUI commands

You typeWhat happens
the history of jazzstart a new topic
/next · /nnext bite-sized lesson
/quiz · /qspaced recall quiz on an earlier lesson
/deeper · /dgo one layer deeper on the current idea
/why <reason>tell the tutor why you're learning this (grounds lessons)
/topic <name>switch to / start another topic
/topicslist saved topics (all resumable)
any questionask the tutor — answered in the topic's own context
/quit · Ctrl+Cexit

Auto-open (Windows Terminal only)

Set "autoLaunch": true in config and the SessionStart hook will open sidequest automatically whenever you start Claude Code — docking into the current window if you're in Windows Terminal, otherwise in its own window.


Supported terminals

TerminalDocked side-by-side?
Windows Terminal✅ via sidequest pair
Warp / PowerShell / cmdSeparate window (no split API)
VS Code integrated terminalSeparate window

Only Windows Terminal exposes a CLI to split itself, so true in-window docking is WT-specific. Everywhere else, sidequest opens its own window beside your agent.


The learning design

The bet: you lose 30–90 seconds dozens of times a day waiting on an agent — enough for one idea. So sidequest is built for micro-learning in the gaps, using techniques from the science of durable learning (the pedagogy is adapted from Matt Pocock's teach skill):

  • Mission grounding — tell it why you're learning a topic (/why) and lessons tie back to that reason.
  • One idea per lesson — a single concrete takeaway; low working-memory load, no dumps.
  • Zone of proximal development — pitched just above what you already know; builds on earlier lessons, never repeats.
  • Retrieval practice + spacing — every few lessons it interleaves a /quiz on an earlier concept. Effortful recall (not re-reading) is what builds long-term retention, so the goal is storage strength, not the fluency illusion.
  • Tight feedback loops — answer a quiz or ask anything and you get immediate, honest correction.
  • Resume, don't restart — per-topic memory spaces practice across many short sittings.

Roadmap

  • Claude Code adapter (hooks → busy/idle)
  • Pluggable provider (GLM / Anthropic / OpenAI)
  • Topic-based bite-sized lessons with persistent context
  • Ink TUI + pair (side-by-side) launcher
  • Retrieval-practice quizzes, mission grounding & tight feedback (the learning method above)
  • Codex CLI adapter (notify in ~/.codex/config.toml)
  • Universal PTY wrapper (sidequest run -- <agent>) for any CLI
  • Full spaced-repetition scheduling (review intervals across sessions)

Development

npm run typecheck     # strict TypeScript, no emit
npm test              # headless suite: core state, hook round-trip, TUI render
npm run test:live     # full pipeline against the live model (needs a key in .env)
npm run doctor        # show config / keys / runtime status

Layout:

src/
  cli.ts            # entry: run | pair | install claude | doctor
  config.ts         # ~/.sidequest config + runtime
  providers/        # Provider interface + glm/anthropic/openai
  core/             # AgentMonitor (busy/idle) + localhost event server
  learning/         # engine, per-topic store, tutor prompts
  tui/App.tsx       # the Ink panel
  install/claude.ts # merges hooks into settings.json
hooks/              # forward.mjs (events) + autostart.mjs (auto-open)

License

MIT

Ver o original no GitHub

Bom para

Relacionados