Review vs Agentic Modes

Agents run in one of two modes depending on the task.

Modes

Mode	Tools Available	Used By
Review (default)	Read, Glob, Grep	`roborev review`, `roborev prompt`
Agentic	Read, Glob, Grep, Edit, Write, Bash	`roborev refine`, `roborev prompt --agentic`

Review Mode

Review mode is read-only - agents can inspect code but cannot make changes. This is the safe default for automated reviews triggered by post-commit hooks.

The agent can:

Read files
Search for patterns
Analyze code structure

The agent cannot:

Edit files
Create files
Run commands

Agentic Mode

Agentic mode allows agents to edit files and run commands. Enable it in three ways:

1. Per-Job

roborev prompt --agentic "Refactor the error handling"
roborev prompt --yolo "Add input validation"   # --yolo is an alias

2. Per-Command

The refine command automatically enables agentic mode:

roborev refine   # Always runs in agentic mode

3. Globally

Set in your global config:

allow_unsafe_agents = true

Then restart the daemon:

roborev daemon restart

Agent-Specific Implementation

Agent	Review Mode	Agentic Mode
Codex	`--sandbox-cmd-allowlist ""`	`--dangerously-bypass-approvals-and-sandbox`
Claude Code	Default behavior	`--dangerously-skip-permissions`
Gemini	Default behavior	`--yolo --allowed-tools`
Copilot	Default behavior	Manual approval required
OpenCode	Default behavior	Auto-approves in non-interactive mode

Security Considerations

When using agentic mode:

The agent can run arbitrary commands on your machine
This includes installing dependencies, running builds, etc.
Safe for your own code on trusted branches
Use isolation for untrusted code (containers, VMs)

See Auto-Fixing Security for detailed guidance.