Sidequest – Aprenda enquanto seu agent trabalha
intincrab/sidequest
Reclaime o tempo morto: aprenda enquanto Claude Code trabalha
Ver no GitHubPainel 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/sidequest1 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
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
- The panel runs a tiny localhost listener (
127.0.0.1:4317by default). - Claude Code hooks (installed into
settings.json) fire a fire-and-forget script on everyUserPromptSubmit/ 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). - While the agent is busy, the panel auto-serves the next bite-sized lesson on your topic.
- Lessons come from a pluggable cheap model — Zhipu GLM, Anthropic Haiku, or an OpenAI mini model.
- 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
pairlayout (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
| Provider | Set SIDEQUEST_PROVIDER | Key env var | Default model | Notes |
|---|---|---|---|---|
| Zhipu GLM | glm | GLM_API_KEY | glm-4.5-flash | OpenAI-compatible; free tier |
| Anthropic | anthropic | ANTHROPIC_API_KEY | claude-haiku-4-5 | — |
| OpenAI | openai | OPENAI_API_KEY | gpt-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 type | What happens |
|---|---|
the history of jazz | start a new topic |
/next · /n | next bite-sized lesson |
/quiz · /q | spaced recall quiz on an earlier lesson |
/deeper · /d | go 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 |
/topics | list saved topics (all resumable) |
| any question | ask the tutor — answered in the topic's own context |
/quit · Ctrl+C | exit |
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
| Terminal | Docked side-by-side? |
|---|---|
| Windows Terminal | ✅ via sidequest pair |
| Warp / PowerShell / cmd | Separate window (no split API) |
| VS Code integrated terminal | Separate 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
/quizon 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 (
notifyin~/.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