Voice-powered code review for human oversight of AI-assisted development.
They have the Golden Arches. We have the Golden Arcs.
They auto-commit AI-generated code without review. We engage in thoughtful voice conversation first.
{counselor} intercepts git commits and engages you in a voice conversation to review code changes before they proceed. Built for developers who use AI coding assistants (like Claude Code, Copilot, Cursor) and want to maintain meaningful human control over what gets committed.
The developer always has final say. Nothing commits without your explicit voice approval. You are the ruler of your codebase — as Prince Akeem would say, you must choose for yourself.
- Hands stay on keyboard — no context switching to review UIs
- Natural conversation — ask questions, get clarification, discuss concerns
- Forced attention — you can't just click "approve" without engaging
- Audit trail — every session is logged with full transcript
Think of it as your royal counselor — the one who actually tells you the truth, not just what you want to hear.
# Install from GitHub
pak::pak("dalyanalytics/counselor")
# Or with devtools
devtools::install_github("dalyanalytics/counselor")Run the interactive setup wizard:
counselor::counselor_setup()This will guide you through getting and configuring your API keys:
| Service | Purpose | Get it at |
|---|---|---|
| Anthropic | Claude for conversation intelligence | console.anthropic.com |
| Deepgram | Speech-to-text | console.deepgram.com |
| Cartesia | Text-to-speech | cartesia.ai |
Keys are stored globally in ~/.Renviron so they work across all your R projects.
Using 1Password? Run counselor_setup(use_1password = TRUE) for secure credential management via 1Password CLI.
Check your configuration anytime:
counselor::api_key_status()In any project where you want voice reviews:
counselor::install_hooks()Now every git commit will trigger a voice review session.
- making a change to test things out
git add .
git commit -m "Add new feature"
# Voice review starts automatically
# Say "approve" to proceed, "reject" to abort, or "goodbye" to exit gracefully# Review currently staged changes
counselor::review_commit()
# Review without voice (for testing)
counselor::review_commit(voice = FALSE)# Start a voice chat (no git context)
counselor::start_session()# Skip for a single commit (the royal court will be disappointed)
COUNSELOR_SKIP=1 git commit -m "Quick fix"During a review session:
| Say | Effect |
|---|---|
| "approve", "looks good", "ship it", "lgtm" | ✅ Approve and proceed with commit |
| "reject", "abort", "cancel", "wait" | ❌ Abort the commit |
| "goodbye", "exit", "bye", "I'm done" | 👋 End session gracefully (commit stays staged) |
| Any question | 💬 AI explains that part of the diff |
The "goodbye" option is perfect for when you need to step away or want to review manually without making a decision yet.
The AI analyzes your staged changes and highlights:
- Security concerns: SQL injection, eval/parse usage, hardcoded credentials
- Shiny-specific issues: Reactive dependencies, renderUI risks
- Risky operations: File deletion, system commands, external downloads
- Sensitive files: .env, credentials, keys being modified
{counselor} includes a comprehensive security module to protect against potentially dangerous operations triggered by voice input:
# Validate voice input against command denylist
counselor::validate_voice_input("please review the auth changes") # OK
counselor::validate_voice_input("run sudo rm -rf /") # Blocked!
# Check if an operation needs confirmation
counselor::check_risky_operation("delete the old test files")
# Returns: list(risky = TRUE, patterns = "delete|remove|rm\\s", ...)
# Request voice confirmation for risky operations
if (counselor::request_voice_confirmation("delete all test files")) {
# User confirmed verbally
}
# Sanitize code before sending to LLM (masks API keys, passwords)
counselor::sanitize_code_for_review('api_key <- "sk-1234567890"')
# Returns: 'api_key = "[REDACTED]"'
# Safe eval wrapper with function allow list
counselor::safe_eval("1 + 1") # OK
counselor::safe_eval("system('ls')") # Blocked!
# Audit security events
counselor::audit_security_logs(days = 30)Blocked patterns include: rm -rf, sudo, system(), eval($()), credential exposure, and more.
Risky patterns that request confirmation: file deletion, git push --force, DROP TABLE, package installations.
All security events are logged to ~/.counselor/security_logs/ for auditing.
Sessions are logged with token usage and cost:
# See your usage
counselor::cost_summary()
# View detailed logs
counselor::read_logs(months = 3)Typical session cost: ~$0.02-0.05
Cartesia offers multiple voices. Find voice IDs at play.cartesia.ai:
counselor::voice_speak("Hello!", voice_id = "your-preferred-voice-id")counselor::review_commit(model = "claude-sonnet-4-20250514")# Check hook status
counselor::hook_status()
# Remove hooks
counselor::remove_hooks()
# Restore backup hook
counselor::remove_hooks(restore_backup = TRUE)- R: Package interface, git integration via {git2r}, conversation via {ellmer}
- Python (via {reticulate}): Voice I/O using Pipecat services
- Deepgram: Nova-2 model for speech-to-text
- Cartesia: Sonic-2 model for text-to-speech
- Claude: Conversation intelligence via {ellmer}
- R >= 4.1.0
- Python >= 3.10
- Working microphone and speakers
- macOS, Linux, or Windows
- Check your microphone permissions
- Ensure your mic is set as default input device
- Try
counselor::test_voice(test_listen = TRUE)
# Check Python config
reticulate::py_config()
# Force reinstall
reticulate::py_install(c("pipecat-ai[cartesia,deepgram]", "pyaudio"))brew install portaudio
pip install pyaudio# Verify installation
counselor::hook_status()
# Check hook is executable
# In terminal: ls -la .git/hooks/pre-commit- Local processing: Audio is processed via API calls, not stored
- No training: Your code/voice is not used to train models
- Logs are local: Session transcripts stay on your machine (
~/.counselor/logs/) - Keys are yours: You control your own API keys
- Security auditing: All blocked commands are logged for review
Prince Akeem didn't want an arranged marriage. He wanted to find someone who would challenge him, question him, and choose him for who he really was.
Your codebase deserves the same consideration. Don't just accept whatever the AI assistant gives you. Engage with it. Question it. Understand it. Then — and only then — commit.
{counselor} is your barbershop moment. The place where real talk happens before you make your move.
MIT
Jasmine Daly (@dalyanalytics)
Built for the 2026 RainbowR conference talk on maintaining human oversight in AI-assisted development.
"Zamunda, Wakanda, Connecticut — I don't know where you from..." — Tracy Morgan, Coming 2 America
Well, this one's from Connecticut. And unlike the fictional kingdoms of Zamunda and Wakanda, we don't have vibranium or royal rose bearers — but we do have {counselor}.
"Just let your Soul Glo!" ✨