Learning-first AI coding assistant for VS Code. Most AI tools write code for you. VibeLearn helps you write it yourself.

Instead of dumping full solutions, VibeLearn asks clarifying questions, hands you hints, names the underlying concept, and only reveals code when you're stuck. You stay in the driver's seat and actually learn while you build.

Autocomplete-style AI helpers are great for shipping. They are not great for learning. If a beginner asks "build me a React login page," a normal assistant emits 80 lines they didn't think about. VibeLearn replies: "What pieces do you think a login page needs? Show me your starting JSX." Then it walks alongside.

You pick how much help you want — from strict (questions only) to full (real answers, with the concept still called out).

• Sidebar chat with a mortarboard icon in the Activity Bar. Built-in welcome, suggested starter prompts, animated "thinking…" indicator, error bubbles, clear-chat button.
• Sidebar action buttons — Project, Learn, and Code action groups let you trigger key commands (Start Session, Dashboard, Reflect, Explain Back, Review Code, etc.) directly from the sidebar without opening the Command Palette.
• Help-level selector in the sidebar header — switch tutoring intensity any time.
• Multiple providers behind a clean interface: OpenAI ✓, Ollama ✓ (local, free), Anthropic / Gemini / OpenRouter (placeholders).
• Fuzzy Ollama model resolution — Llama 3.2, llama 3.2, LLAMA3.2, even llma3.2 all resolve to llama3.2:latest.
• Code review command — select code, right-click → VibeLearn: Review Selected Code. Reply structured as: what it does / what's good / what could be improved / one hint to try next.
• Prompt rewriter — turn "build me a todo app" into a learning-focused prompt you can paste into any AI tool. Local templates, no API call.
• Model picker — dropdown of installed Ollama models, or curated suggestions for cloud providers.
• Secure API keys — stored in OS keychain via VS Code SecretStorage. Never in settings.json, never logged.

Prerequisites: Node.js 18+, VS Code 1.85+.

npm install
npm run compile
npm test          # run unit tests

In VS Code:

1. Open this folder.
2. Press F5 (or Fn+F5 on Mac) → Extension Development Host opens.
3. Click the mortarboard icon in the left Activity Bar.
4. Configure a provider (see below) and start chatting.

If you haven't configured a provider yet, the sidebar shows a Welcome card with "Choose Model" and "Set API Key" buttons. The action buttons and chat are hidden until setup is complete.

• OpenAI — fastest path. Get a key from platform.openai.com, add billing, then run VibeLearn: Set API Key.
• Ollama — free + local. Install from ollama.com (or brew install ollama), then ollama serve and ollama pull llama3.2. No key needed. See the Ollama section below.

Open the Command Palette with Cmd+Shift+P (Mac) or Ctrl+Shift+P (Win/Linux) and search VibeLearn:

Cmd+, → search vibelearn.

• strict — only clarifying questions and concept hints. Refuses code.
• guided (default) — hints first, pseudocode if useful, full code only on explicit request.
• assist — quick concept hint, then partial code with explanation.
• full — full solutions, but still names the underlying concept.

Each client implements the same AIClient interface (src/ai/types.ts). Adding a provider = drop a new file in src/ai/, add a case to the factory in src/ai/index.ts. The teaching system prompt and chat history wiring stay shared.

1. Cmd+Shift+P → VibeLearn: Set API Key → pick openai → paste key.
2. Cmd+, → confirm:
   • vibelearn.provider = openai
   • vibelearn.model = e.g. gpt-4o-mini
3. Open the sidebar (mortarboard icon). Ask a question.

Free, private, offline. Requires installing Ollama once.

1. Install: download from ollama.com, or brew install ollama.
2. Start the server (leave running):

   ollama serve

   Listens on http://localhost:11434.
3. Pull a model. Start small:

   ollama pull llama3.2          # ~2 GB, general
   ollama pull qwen2.5-coder:7b  # ~4 GB, code-focused

   List installed: ollama list.

1. Cmd+, → search vibelearn:
   • vibelearn.provider = ollama
   • vibelearn.model = the tag from ollama list (e.g. llama3.2). Capitalization and spaces are forgiven.
2. No API key step. Open the sidebar and chat.

First reply may be slow while the model loads into RAM; subsequent replies are fast.

A Learning Session ties together a project roadmap, attempt-first guidance, and your learning profile into a single focused experience.

Start a session: VibeLearn: Start Learning Session → enter a project idea → VibeLearn generates a roadmap, extracts milestones, and saves the session locally.

While a session is active, every chat message includes your current milestone as context so the AI keeps guidance focused on what you're building right now.

Track progress: VibeLearn: Show Learning Session opens a Markdown document with a checklist of milestones and your current position.

Advance: VibeLearn: Complete Current Milestone marks the current milestone done and moves to the next. When the final milestone is completed, you get a congratulatory message.

One active session at a time. Sessions are stored in VS Code globalState — local only, never synced.

VibeLearn tracks what you encounter across chat sessions and uses it to personalise teaching.

After each AI reply, a local heuristic scans the response for:

• Concepts — e.g. array, loop, async, recursion, authentication. Added to Concepts Seen.
• Strength signals — phrases like "you understand", "good job", "that's correct". Logged as a strength.
• Struggle signals — phrases like "common mistake", "watch out", "tricky part". Logged as a struggle.

The profile is injected as optional context into the system prompt so the AI can avoid re-explaining things you already know and give extra care to areas where you've struggled.

Privacy: The profile is stored entirely in VS Code globalState on your local machine. It is never sent anywhere except as part of the prompt to your configured AI provider (OpenAI or Ollama). It is never logged or synced.

Use VibeLearn: Show Learning Profile to inspect it and VibeLearn: Clear Learning Profile to wipe it.

API keys live in VS Code SecretStorage, which delegates to the OS keychain (macOS Keychain, Windows Credential Manager, libsecret on Linux).

• Not in settings.json.
• Not in any extension log or telemetry.
• Not echoed back in error messages — error paths read only the provider's error.message field.
• Not persisted in chat history sent to the model — only your messages and assistant replies are.
• Per-provider slots, so storing a Claude key doesn't overwrite your OpenAI one.

You can clear a key any time with VibeLearn: Clear API Key. Ollama uses no key at all.

.
├── src/
│   ├── extension.ts             # activate, command registration
│   ├── chatViewProvider.ts      # sidebar webview UI + chat orchestration
│   ├── secrets.ts               # SecretStorage helpers
│   ├── modelPicker.ts           # provider-aware model picker
│   ├── promptRewrite.ts         # local prompt-rewrite templates
│   ├── promptRewrite.test.ts
│   ├── learningProfile.ts       # local learning profile (globalState)
│   ├── learningProfile.test.ts
│   ├── learningSession.ts       # learning session / project mode (globalState)
│   ├── learningSession.test.ts
│   ├── workspaceContext.ts      # workspace structure reader and formatter
│   ├── workspaceContext.test.ts
│   ├── dashboard.ts             # learning dashboard markdown builder
│   ├── dashboard.test.ts
│   └── ai/
│       ├── types.ts             # AIClient, AIRequest, AIError, Provider, ChatMessage
│       ├── promptBuilder.ts     # teaching system prompt per helpLevel
│       ├── promptBuilder.test.ts
│       ├── roadmapPrompt.ts     # roadmap message builder
│       ├── roadmapPrompt.test.ts
│       ├── attemptDetector.ts   # heuristic: has the user shown an attempt?
│       ├── attemptDetector.test.ts
│       ├── reflectionPrompt.ts  # reflection question builder
│       ├── reflectionPrompt.test.ts
│       ├── explainBackPrompt.ts # explain-back review builder
│       ├── explainBackPrompt.test.ts
│       ├── modelMatch.ts        # fuzzy Ollama model resolution
│       ├── modelMatch.test.ts
│       ├── notImplementedClient.ts
│       ├── openaiClient.ts      # OpenAI (implemented)
│       ├── ollamaClient.ts      # Ollama (implemented)
│       ├── anthropicClient.ts   # placeholder
│       ├── geminiClient.ts      # placeholder
│       ├── openrouterClient.ts  # placeholder
│       ├── factory.test.ts
│       └── index.ts             # provider factory
├── media/vibelearn.svg          # activity bar icon
├── docs/
│   ├── project-overview.md
│   └── teaching-rules.md
├── package.json
├── tsconfig.json
└── README.md

• Project Overview — what VibeLearn is and why.
• Teaching Rules — how the assistant should behave.

MIT — see LICENSE.