Code Review Workflow

What It Does

ReviewOrchestrator manages a two-agent review-fix iteration cycle. A reviewer agent examines a diff and produces structured ReviewFeedback with comments. If the feedback contains errors, an author agent receives the comments and applies fixes. The cycle repeats until the reviewer approves or max_rounds is reached.

CLI

chimera review --diff changes.patch --model glm-5 --max-rounds 3

Python API

from chimera.review import ReviewOrchestrator, ReviewFeedback
from chimera.core.agent import Agent
from chimera.providers.factory import create_provider

provider = create_provider(model="glm-5")
reviewer = Agent(provider=provider)
author = Agent(provider=provider)

orchestrator = ReviewOrchestrator(max_rounds=3)
approved = orchestrator.run(
    diff=open("changes.patch").read(),
    reviewer=reviewer,
    author=author,
    env=None,
)

print(f"Approved: {approved}")
print(f"Rounds: {orchestrator.current_round}")
print(f"Total comments: {orchestrator.total_comments}")

Key Classes

ReviewOrchestrator

class ReviewOrchestrator:
    def __init__(self, max_rounds: int = 3) -> None
    def run(self, diff: str, reviewer: Agent, author: Agent, env: Environment | None = None) -> bool
    def add_review(self, feedback: ReviewFeedback) -> ReviewRound
    def mark_fixed(self) -> None
    def needs_another_round(self) -> bool

Properties: max_rounds, rounds (list of ReviewRound), current_round, is_approved, is_complete, total_comments.

ReviewFeedback

Dataclass with fields: comments (list of ReviewComment), approved (bool), summary (str).

Properties: has_critical, has_errors, comment_count, files_reviewed.

Methods: by_severity(severity), by_file(file), parse_from_text(text) (static, parses [SEVERITY] file:line: message format).

ReviewComment

Dataclass with fields: file, line, severity (Severity enum), message, suggestion.

Severity

Enum: INFO, SUGGESTION, WARNING, ERROR, CRITICAL.

Integration

• The orchestrator is a specialised two-agent composition (reviewer + author iteration); drop it into a Pipeline if you want to chain it with planning or testing stages.
• ReviewFeedback.parse_from_text(text) recognises the [SEVERITY] file:line: message line format the reviewer agent is prompted to emit. Override the prompt for a different output shape and supply a custom parser.
• For multi-perspective reviews (security / performance / maintainability / etc.), see chimera.review.perspective and the review_server MCP server.
• The CLI chimera review --diff <patch> is a thin wrapper around ReviewOrchestrator.run().

Import

from chimera.review import (
    ReviewComment,
    ReviewFeedback,
    ReviewOrchestrator,
    Severity,
)