AI Supervisor for Claude Code: Autonomous Approval with Safety Rules

Running Claude Code for extended sessions means dealing with a constant stream of approval prompts. Auto-Yes handles the simple cases, but what about sessions that need judgment — approving safe file edits while blocking destructive commands, answering the agent's questions, and knowing when to escalate to a human? That's what Remocode's AI Supervisor does.

What the AI Supervisor Actually Does

The Supervisor is a lightweight AI layer that sits between Claude Code and your keyboard. It reads the terminal, understands what Claude Code is asking, and makes a decision — all within a 2-second scan cycle.

The Decision Categories

Every prompt from Claude Code gets classified into one of five responses:

• ●Approve — Safe action, send the appropriate keystroke to confirm
• ●Reject — Dangerous action, send Escape to cancel
• ●Answer — Claude Code asked a question, type a response and press Enter
• ●Escalate — Uncertain, show the decision in the AI panel for human review
• ●Navigate menu — Claude Code shows a numbered menu, select the right option using arrow keys

How It Reads the Terminal

Every 2 seconds, the Supervisor captures the last 20 lines of terminal output. It sends this context, along with your project brief, to a configured AI model. The model returns a structured JSON response with an action and reasoning.

Writing an Effective Project Brief

The project brief is the Supervisor's instruction manual. It's free-form natural language — write it however makes sense for your project.

Example Brief

Building a Next.js e-commerce dashboard.
Stack: Next.js 14, TypeScript, Prisma, PostgreSQL, Tailwind CSS.

Approve: File edits, component creation, test runs, npm install, migrations.
Reject: Any rm -rf command, sudo usage, production database operations.
Answer: Use PostgreSQL for all database questions. Prefer server components.
Escalate: Anything involving environment variables or API keys.

Brief Writing Tips

• ●Be specific about your stack — helps the Supervisor answer technical questions
• ●List explicit approve/reject categories — reduces ambiguity
• ●Include project conventions — naming patterns, folder structure preferences
• ●Mention sensitive areas — files or directories that should never be modified

Autonomy Levels

The Supervisor supports three autonomy levels that control how aggressively it acts:

Conservative

Approves only actions explicitly matching your brief. Everything else gets escalated. Best for production codebases or unfamiliar projects where you want maximum oversight.

Balanced

Approves routine operations (file edits, test runs, installs) and rejects known-dangerous patterns. Escalates edge cases. The default for most workflows.

Aggressive

Approves nearly everything except hard-blocked patterns (rm -rf, sudo, DROP TABLE). Minimal escalation. Best for greenfield projects or trusted refactoring tasks.

Safety Rules: What Gets Blocked

Regardless of autonomy level, the Supervisor has hard-coded safety patterns that always trigger rejection:

• ●`rm -rf` — Recursive forced deletion
• ●`sudo` — Privilege escalation
• ●`DROP TABLE` / `DROP DATABASE` — Destructive SQL operations
• ●`chmod 777` — Overly permissive file permissions
• ●Force push to main — git push --force origin main

These are always rejected with a clear reason shown in the decision log.

The Decision Log

Every Supervisor decision is logged with full context:

• ●Timestamp of the decision
• ●Terminal content that triggered it
• ●Action taken (approve, reject, answer, escalate)
• ●Reasoning explaining why the AI made that choice
• ●Model used for the decision

The log appears in real-time in Remocode's AI panel. You'll see "Thinking..." while the model processes, then the result with color-coded status: green for approved, red for rejected, yellow for escalated.

Choosing a Model for the Supervisor

The Supervisor works with any model configured in Remocode. Since the decisions are relatively simple (classify a terminal prompt into approve/reject/answer/escalate), cheap and fast models work best:

• ●GPT-5-nano — Fast, cheap, accurate for simple decisions
• ●Haiku — Anthropic's lightweight model, great for quick classification
• ●Local models via Ollama — Zero cost, runs entirely on your machine

Expensive models like GPT-5 or Opus are overkill for supervisor duties. Save those for your actual coding agent.

Auto-Yes vs. Supervisor: When to Switch

The two features are mutually exclusive — enabling one disables the other on that pane. The rule of thumb:

• ●Auto-Yes: You trust the task completely, want zero cost
• ●Supervisor: You want autonomous operation with guardrails

You can mix them across panes. Run Auto-Yes on a scaffolding pane and the Supervisor on a refactoring pane in the same Remocode window.

Conclusion

The AI Supervisor turns Claude Code into a truly autonomous coding agent — one that respects your project context, blocks dangerous operations, answers questions intelligently, and logs every decision with reasoning. Write a brief, pick an autonomy level, and let it run.