multi-model-review

Multi-Model Code Review

Review code changes by broadcasting to multiple AI models in parallel. One call builds context once and fans out — faster and more consistent than sequential reviews.

Quick Start

# Review uncommitted changes with all CLIs
broadcast(
  prompt="Review this diff for bugs, security issues, and improvements",
  cd="/path/to/project",
  review_uncommitted=True
)

# Review against a branch (e.g. PR review)
broadcast(
  prompt="Review these changes for correctness and style",
  cd="/path/to/project",
  review_base="main"
)

# Review a specific commit
broadcast(
  prompt="Review this commit",
  cd="/path/to/project",
  review_commit="abc1234"
)

Choosing Targets

# All three CLIs (default)
broadcast(prompt="...", cd="...", review_uncommitted=True)

# Specific targets
broadcast(prompt="...", cd="...", targets=["codex", "gemini"], review_uncommitted=True)

Recommendation: Use targets=["codex", "gemini"] for code review. Two perspectives are usually sufficient and faster than three.

Review Modes

Mode	Parameter	Use Case
Uncommitted	review_uncommitted=True	Review working tree changes before commit
Branch diff	review_base="main"	Review all changes on a feature branch
Commit	review_commit="abc1234"	Review a specific commit (7-40 hex SHA)
File list	files=["src/foo.py"]	Review specific files (relative paths only)

Modes are mutually exclusive. files can combine with review modes to add extra context.

Per-Target Configuration

broadcast(
  prompt="Review this diff",
  cd="/path/to/project",
  review_uncommitted=True,
  models={"gemini": "gemini-3.1-pro-preview"},
  overrides={
    "codex": {"reasoning_effort": "high"},
    "claude": {"effort": "high"}
  }
)

Interpreting Results

The response contains one result per target. Compare them:

• Agreement on an issue = high confidence finding
• Unique findings = worth investigating
• Contradictions = use your judgment, or ask a follow-up

Best Practices

• Always use broadcast instead of calling codex + gemini + claude separately — it builds context once
• For large diffs, add system_prompt="Focus on critical bugs and security issues only" to keep output focused
• Save session_id from each target's result if you want follow-up questions per model
• Check success field for each target — one target failing doesn't block others