pi-learn-btw

Learn English By The Way, while coding with AI.
Zero-cognitive-load, multi-language expression improvement for Pi Coding Agent.

💡 Philosophy: Learn By The Way

Existing AI language tools either:

1. Pollute conversation context: Inject grammar lectures directly into chat history, consuming expensive reasoning tokens on every subsequent turn.
2. Interrupt your flow: Force you into a "language lesson" mindset while you are solving engineering problems.
3. Scroll away: In terminal coding agents, tool calls, bash logs, and thinking blocks push your original prompt and any feedback off the screen.

pi-learn-btw solves this with a clean, decoupled architecture:

• 📌 Pinned belowEditor: Pinned directly below your input box. It never scrolls away.
• ⚡ Zero Friction: Uses models you have already configured in Pi. Your primary task runs immediately, while your expression is refined in parallel.
• 🎨 Git Code Diff Theme: Soft dark-red (removed words) and soft dark-green (improved words) in a clean 2-line stacked diff.
• ⌨️ Instant Toggle (Alt+E): Cycle seamlessly between [DIFF] ➔ [IMPROVED] ➔ [ORIGINAL] with clear mode badges.
• 🌐 Multi-Language Ready: Learn English (recommended), Spanish, French, German, Italian, Portuguese, or any space-delimited language. (Note: Non-spaced languages like Chinese and Japanese are not currently supported by word-level diffing).
• 📊 Quantified Progress (/learn-btw stats): Tracks your modification ratio and error reduction percentage over time.

📦 Installation

Install directly via Pi's package manager:

# Install from npm (recommended)
pi install npm:pi-learn-btw

# Or install from GitHub
pi install git:github.com/chxsong/pi-learn-btw

On first startup, pi-learn-btw greets you with a quick setup prompt to confirm or customize your coach model and target language.

🤖 Model Configuration

pi-learn-btw uses the models you have already configured in Pi:

• Default: Automatically uses your active Pi session model (0 additional setup required).
• Custom Coach: Want to use a cheaper/faster model (like gpt-4o-mini, claude-3-5-haiku, or gemini-2.5-flash) for expression coaching? Run /learn-btw config and select from any model configured in your Pi environment.

You can also override the coach provider and model via environment variables:

• PI_LEARN_BTW_PROVIDER: e.g. openai, anthropic, google
• PI_LEARN_BTW_MODEL: e.g. gpt-4o-mini, claude-3-5-haiku, gemini-2.5-flash

🎮 Commands: /learn-btw (Alias: /btw)

/learn-btw          -> Toggle on / off (quick one-shot toggle)
/learn-btw stats    -> View learning analytics (prompts analyzed, improvement rate, top words)
/learn-btw config   -> Open the interactive configuration wizard

Keyboard Shortcut

• Alt+E: Cycle between [DIFF] ➔ [IMPROVED] ➔ [ORIGINAL] views.

⚙️ Interactive Configuration (/learn-btw config)

Running /learn-btw config opens an interactive TUI menu:

1. Target Language: English (recommended), Spanish, French, German, Italian, Portuguese, or Custom (space-delimited).
2. Send Polished Prompt to AI: Disabled (0ms latency, sends what you typed) or Enabled (waits for polish, sends improved prompt to AI).
3. Select Model: Choose from the models you have already configured in Pi, or enter a custom provider/model_id.
4. Default View: diff, improved, or original.
5. Clear History Data: Reset local analytics records.

Configuration is persisted in ~/.pi/learn-btw/config.json.

🧪 Testing & Verification

Run the automated test suite and type checking:

npm test       # Run 12 unit tests
npm run check  # TypeScript strict type checking

All 12 unit tests cover:

• Word-level LCS diffing without word concatenation
• Soft code-diff ANSI terminal color rendering
• Case-insensitivity and punctuation-ignoring gatekeeper
• JSON parsing and error resilience
• History persistence and metrics calculation

📄 License

MIT © chxsong