Auto-Fixing with Refine

Use refine to automatically address failed reviews on your branch.

roborev refine                       # Fix failed reviews using default agent
roborev refine --agent claude-code   # Use specific agent for addressing findings
roborev refine --max-iterations 5    # Limit fix attempts
roborev refine --quiet               # Show elapsed time instead of agent output
roborev refine --reasoning thorough  # Use thorough reasoning (slower, more accurate)
roborev refine --since abc123        # Refine commits since a specific commit
roborev refine --since HEAD~3        # Refine the last 3 commits (works on main branch)

How It Works

flowchart TD
  Start([refine]) --> Find[Find oldest failed review]
  Find --> A{Found?}
  A -->|No| Branch[Run whole-branch review]
  Branch --> C{Pass?}
  C -->|Yes| Done([Done - all reviews pass])
  C -->|No| Fix
  A -->|Yes| Fix[AI agent addresses findings]
  Fix --> Commit[Commit fix]
  Commit --> Wait[Wait for re-review]
  Wait --> Pass{Pass?}
  Pass -->|Yes| Find
  Pass -->|No| MaxCheck{Max iterations?}
  MaxCheck -->|No| Fix
  MaxCheck -->|Yes| Fail([Exit - max reached])

Processing Order

The loop processes commits oldest-first
When a fix fails review, it keeps addressing that fix until it passes
Passing reviews are automatically marked as addressed
When no individual failed reviews remain, a whole-branch review runs
If that passes, the refiner exits successfully

Options

Flag	Description
`--agent <name>`	Agent to use for fixes
`--max-iterations <n>`	Maximum fix attempts per commit (default: 3)
`--quiet`	Show elapsed time instead of agent output
`--reasoning <level>`	Reasoning depth: fast, standard, thorough
`--since <commit>`	Refine commits since a specific commit

Requirements

Clean working tree: No uncommitted changes
Feature branch: By default, compares against main
Use --since: To refine specific commits on any branch, including main

Security Considerations

The refine command runs AI agents without sandboxing so they can install dependencies, run builds, and execute tests. This is safe for your own code, but use caution with untrusted sources:

Scenario	Risk Level	Recommendation
Your own branches	Low	Safe to run directly
PRs from trusted contributors	Low-Medium	Review changes before refining
PRs from strangers / untrusted code	High	Use isolation (see below)

Isolation Options for Untrusted Code

Run in a container or VM
Use a low-privilege user account
Use a disposable cloud instance