Keywords: code review, parallel, multi-model, consensus, disagreement, prioritized findings

Run three independent code reviews in parallel using Anthropic, OpenAI, and Google models via the software-code-reviewer agent, then synthesize a unified prioritized findings report highlighting reviewer agreement and disagreement

Keywords: code review · parallel · multi-model · consensus · disagreement · prioritized findings

Usage

To run this workflow, start your prompt with ~code-review-parallel followed by your goal, or select it with the /workflow command.

When to use

when you want independent code reviews from multiple LLM families to reduce blind spots

when you want a synthesized report that highlights where reviewers agree and disagree

When not to use

when a single-model code review is sufficient

when the task is implementation, refactoring, or test writing rather than review

Configuration

Property	Value
Goal	What code should be reviewed? (e.g., a file path, package name, or description of the change)
Referenced agents	`software-code-reviewer`

Pipeline

This workflow fans out to three parallel code reviews of the same target, each using the software-code-reviewer agent but forced to a different model provider (Anthropic, OpenAI, Google). After all three complete, a synthesis node merges the reviews into a single numbered, prioritized findings list that calls out agreement and disagreement across reviewers. The workflow ends immediately after synthesis so it can be composed by parent workflows without an extra human gate; callers that need human approval should add their own review step after this workflow.

`synthesize-prompt`

You have received three independent code reviews of the same code. Retrieve the output of each review node using the available tools.

Produce a single unified report with these sections:

Summary — a brief overview of the code under review and the overall assessment.

Prioritized Findings — a numbered list of findings ordered by severity (critical → high → medium → low → informational). For each finding:

Title: concise name

Severity: critical / high / medium / low / informational

Agreement: note which reviewers flagged this (e.g., "All three", "A + B", "C only")

Description: what the issue is and why it matters

Recommendation: concrete fix or improvement

Consensus & Disagreement — highlight:

Areas where all three reviewers agreed

Areas where reviewers disagreed or only one reviewer flagged an issue, with brief reasoning about which perspective seems strongest

Overall Recommendation — a single clear recommendation (approve, approve with changes, or request revisions) based on the weight of evidence across all three reviews.

Before finalizing, validate each finding against the actual code. Pay particular attention to findings flagged by only one reviewer — these are the most likely to be either a genuine insight the others missed or a hallucinated issue. Read the relevant code to confirm the finding is real. Downgrade or drop any finding you cannot substantiate, and note when a single-reviewer finding was confirmed or rejected by your own inspection.

Be concise. Do not reproduce the full text of each review — synthesize and deduplicate.

Write the final report to .stencila/reviews/ as a Markdown file. Derive the filename from the review target using kebab-case (e.g., .stencila/reviews/codec-markdown-review.md). Create the directory if it does not exist. If a file with that name already exists, choose another name.

Then output the same report in full as your final message. The final message is the canonical workflow output for composition by parent workflows. Do not respond with only a file path, success message, or summary. After the report, add a final line indicating where the copy was saved.

This page was generated from .stencila/workflows/code-review-parallel/WORKFLOW.md.