AI CodeSensei
Learn any codebase with AI, then test your knowledge through a live, voice-driven conversation.
Built with ❤ by sharadcodes · g-savitha · iamnabina
Overview
AI CodeSensei helps you understand a codebase, then tests your knowledge through a live, voice-driven Knowledge Check session. It uses ACP agents (Codex, Devin, etc.) to analyze your repository, then conducts a real-time conversation where it asks questions about the code, opens the relevant files, and highlights the exact lines it's asking about.
Features
Code Tutor Guide
Generates a comprehensive AI CodeSensei.md guide by reading your repository with your selected ACP agent. Choose from three depths:
- Quick Overview (~5 min) — purpose, stack, entry points, architecture map
- Guided Walkthrough (~10 min) — major modules, data/control flow, conventions
- Deep Dive (15+ min) — architecture, abstractions, workflow tracing, risks
Knowledge Check
A live, voice-driven interview about your codebase:
- Discovers ACP agents (Codex, Devin, OpenCode, Gemini CLI, etc.) from the Agent Client Protocol registry
- Analyzes your codebase — the agent reads your repo and produces a structured learning summary with key files and suggested topics
- Conducts a voice session — speech-to-text captures your answers, an LLM evaluates them, and text-to-speech responds
- Auto-opens files — as it asks about specific code, it opens the file and highlights the exact line range
- Adaptive difficulty — adjusts question difficulty based on your responses
Requirements
- VS Code 1.90+
- An ACP agent installed (e.g.
npx -y @agentclientprotocol/codex-acp)
- An STT API key — OpenRouter or any OpenAI-compatible endpoint
- A chat API key — OpenRouter or any OpenAI-compatible endpoint
- A TTS endpoint — Kokoro FastAPI (local, free) or any OpenAI-compatible TTS
- PortAudio — bundled for Windows; macOS/Linux may need
portaudio installed
Quick Start
- Install AI CodeSensei from the VS Code Marketplace
- Open the workspace you want to learn
- Click the AI CodeSensei icon in the Activity Bar
- Click ↻ Agents to discover available ACP agents
- Select an agent (check icon)
- Click Start session to begin a Knowledge Check
- Speak naturally — the evaluator listens, responds, and asks follow-up questions
Configuration
All settings live under the AI CodeSensei.* namespace. Open Settings (Ctrl+,) and search for AI CodeSensei.
Speech-to-Text (STT)
| Setting |
Default |
Description |
AI CodeSensei.stt.baseUrl |
https://openrouter.ai/api/v1 |
OpenAI-compatible STT endpoint |
AI CodeSensei.stt.model |
mistralai/voxtral-mini-transcribe |
STT model slug |
AI CodeSensei.stt.apiKey |
— |
API key. Falls back to OPENROUTER_API_KEY, then OPENAI_API_KEY |
AI CodeSensei.stt.path |
/audio/transcriptions |
Endpoint path |
AI CodeSensei.stt.language |
en |
ISO-639-1 language code. Empty = auto-detect |
Text-to-Speech (TTS)
| Setting |
Default |
Description |
AI CodeSensei.tts.baseUrl |
http://localhost:8881/v1 |
OpenAI-compatible TTS endpoint (Kokoro FastAPI by default) |
AI CodeSensei.tts.model |
tts-1 |
TTS model |
AI CodeSensei.tts.apiKey |
not-needed |
API key (Kokoro doesn't require one) |
AI CodeSensei.tts.voice |
af_heart |
Voice name (Kokoro: af_heart, af_bella, af_nova, etc.) |
AI CodeSensei.tts.path |
/audio/speech |
Endpoint path |
AI CodeSensei.tts.responseFormat |
wav |
Audio format: wav, flac, ogg, mp3, opus |
Chat (LLM)
| Setting |
Default |
Description |
AI CodeSensei.chat.baseUrl |
https://openrouter.ai/api/v1 |
OpenAI-compatible chat endpoint |
AI CodeSensei.chat.model |
openai/gpt-5.6 |
Chat model for Knowledge Check orchestration |
AI CodeSensei.chat.apiKey |
— |
API key. Falls back to OPENROUTER_API_KEY, then OPENAI_API_KEY |
AI CodeSensei.chat.path |
/chat/completions |
Endpoint path |
Audio
| Setting |
Default |
Description |
AI CodeSensei.audio.inputDeviceId |
-1 |
PortAudio device ID. -1 = system default. Use the in-app dropdown to pick. |
AI CodeSensei.audio.silenceSeconds |
2 |
Seconds of silence before ending a speech segment |
AI CodeSensei.audio.beepEnabled |
true |
Play a beep when it's your turn to speak |
ACP Agents
| Setting |
Default |
Description |
AI CodeSensei.acp.selectedAgentId |
— |
Pre-selected agent ID (codex, devin, etc.) |
AI CodeSensei.acp.agentConfigs |
{} |
Per-agent config map (model, reasoning effort, sandbox, etc.) |
AI CodeSensei.acp.contextPrompt |
(see default) |
Prompt sent to the agent for codebase analysis |
AI CodeSensei.acp.customAgents |
[] |
Additional stdio-based ACP agents |
Interview
| Setting |
Default |
Description |
AI CodeSensei.interview.maxQuestions |
0 |
Max questions (0 = unlimited) |
AI CodeSensei.interview.difficulty |
adaptive |
adaptive, junior, mid, senior, staff |
AI CodeSensei.tutor.explanationMode |
guided |
Guide depth: quick, guided, deep |
Commands
| Command |
Description |
AI CodeSensei: Start Knowledge Check |
Begin a voice-driven interview session |
AI CodeSensei: Stop Active Operation |
Stop the current session or guide generation |
AI CodeSensei: Generate Codebase Guide |
Create a AI CodeSensei.md learning guide |
AI CodeSensei: Refresh ACP Agents |
Re-scan for available agents |
AI CodeSensei: Test Microphone |
Verify mic capture is working |
AI CodeSensei: Test Speaker (TTS) |
Verify TTS playback is working |
AI CodeSensei: Clear Cached Session |
Delete cached analysis and start fresh |
AI CodeSensei: Show Logs |
Open the AI CodeSensei output channel |
Architecture
- ACP layer — discovers and launches agents via the standard Agent Client Protocol registry. Any agent that publishes an
agent.json works automatically.
- Chained voice pipeline — PortAudio mic capture → VAD → STT → LLM chat → TTS → webview playback. All components are swappable behind their interfaces.
- Source policy — the Code Tutor guide generator creates a curated, read-only analysis workspace with only permitted files, preventing the agent from accessing secrets or irrelevant files.
Build VSIX artifacts
.github/workflows/publish-marketplace.yml is triggered manually via workflow_dispatch (Actions tab → Run workflow). It validates the extension, reads the version from package.json, builds separate Apple Silicon macOS, Intel macOS, x64 Linux, and x64 Windows VSIX packages, verifies their native architectures and runtime loading, and uploads the VSIX files as downloadable workflow artifacts.
Windows ARM64 and Linux ARM64 are not currently built because naudiodon2 2.5.0 bundles x64-only PortAudio libraries for those platforms.
Bump the version in package.json on the target branch before triggering the workflow.
License
MIT
| |
