Supervisor for Gemini CLI: Smart Auto-Approval with Project Briefs

Google's Gemini CLI is a capable coding agent, but like all terminal-based AI tools, it interrupts your workflow with approval prompts. Remocode's AI Supervisor adds an intelligent layer between Gemini and your keyboard — one that understands your project context and makes smart approval decisions automatically.

Why Smart Approval Beats Blind Approval

Auto-Yes clicks "Yes" on everything. That's fine for simple, trusted tasks. But Gemini CLI sessions often involve nuance:

• ●Gemini might propose restructuring your directory layout — do you want that?
• ●It might need to know which testing framework you prefer — can it get an answer without you?
• ●It might suggest installing a package with known vulnerabilities — should that be blocked?

The AI Supervisor handles all of these by reading a project brief you write in natural language.

Writing Project Briefs for Gemini CLI Sessions

The brief is the single most important piece of configuration. A well-written brief dramatically reduces escalations and incorrect decisions.

Anatomy of a Good Brief

Project: Mobile app backend (Kotlin, Spring Boot, PostgreSQL).
Deploying to Google Cloud Run.

APPROVE:
- All source file edits in src/main/ and src/test/
- Running ./gradlew build, ./gradlew test
- Installing dependencies via build.gradle changes
- Creating new files in the project structure

REJECT:
- Any rm -rf or recursive delete commands
- sudo or privilege escalation
- Direct database operations (we use migrations only)
- Modifying Dockerfile or cloudbuild.yaml

ANSWER:
- Database: PostgreSQL 16 with Flyway migrations
- Testing: JUnit 5 with Mockito
- API style: REST with OpenAPI 3.0 spec
- Auth: Firebase Auth, JWT tokens

ESCALATE:
- Changes to application.yml or application-prod.yml
- New external API integrations
- Dependency version bumps

Why Explicit Sections Help

The Supervisor sends your brief to the AI model along with the terminal content. Clear APPROVE/REJECT/ANSWER/ESCALATE sections make the model's job easier and produce more predictable decisions. You can write your brief in any format, but explicit categories reduce ambiguity.

How the Supervisor Processes Gemini CLI Prompts

The Scan Cycle

• ●Every 2 seconds, the Supervisor reads the last 20 lines of terminal output
• ●If the content hasn't changed since the last scan, it does nothing (zero AI cost)
• ●If a new prompt is detected, it sends the content and brief to the configured AI model
• ●The model returns a JSON response: action type, the response text, and reasoning
• ●The Supervisor executes the action via keystroke injection

Handling Different Prompt Types

Yes/No confirmations: The Supervisor sends Enter to approve or Escape to reject.

Numbered menus: Gemini CLI sometimes presents multiple options. The Supervisor selects the appropriate option based on your brief and the context.

Free-text questions: When Gemini asks a question like "Which database driver should I use?", the Supervisor types a response based on your brief's ANSWER section and presses Enter.

Uncertain situations: When the Supervisor can't confidently decide, it escalates to the AI panel with its reasoning, letting you make the call.

Autonomy Levels Explained

Conservative

Only approves actions that explicitly match your brief. Everything else is escalated. This level generates more escalations but gives you maximum control. Ideal for:

• ●Production codebases
• ●Unfamiliar projects you're still learning
• ●Regulated environments

Balanced

Approves routine development actions (file edits, test runs, builds) even if not explicitly listed. Rejects known-dangerous patterns. Escalates true edge cases. The default for most developers.

Aggressive

Approves almost everything, only blocking hard-coded safety patterns (rm -rf, sudo, DROP TABLE). Minimal escalation. Use this for:

• ●Greenfield projects with low risk
• ●Sandboxed or containerized environments
• ●Rapid prototyping sessions

Cost Efficiency With Cheap Models

The Supervisor's decisions are simple classification tasks. You don't need a frontier model:

| Model | Cost Per Decision | Speed | Recommendation | |-------|------------------|-------|----------------| | GPT-5-nano | ~$0.0001 | Fast | Best value | | Haiku | ~$0.0001 | Fast | Excellent alternative | | Ollama (local) | $0 | Varies | Best for privacy |

A typical Gemini CLI session generates 10-30 supervisor decisions. At GPT-5-nano pricing, that's less than a cent for the entire session.

Monitoring Decisions in Real-Time

The AI panel in Remocode shows every Supervisor decision as it happens:

• ●Yellow "Thinking..." — The model is processing a detected prompt
• ●Green "Done" — Decision made, action executed, reasoning displayed
• ●Red "Error" — AI call failed (network issue, rate limit, etc.)

This real-time feed means you can glance at the AI panel occasionally without interrupting your own work. If something looks wrong, you can disable the Supervisor and take over manually.

Conclusion

The AI Supervisor transforms Gemini CLI from an interactive tool that demands constant attention into an autonomous agent that respects your project's rules. Write a clear project brief, choose your autonomy level, pick a cheap model, and let the Supervisor handle the rest.