How Code Review Confidence Scoring Works in Claude Code

Claude Code's code review confidence scoring assigns a 0-100 numeric value to each detected issue, filtering out any results below a configurable threshold (default 80) to ensure only high-signal feedback reaches developers.

The Code Review plugin in the anthropics/claude-code repository employs a multi-agent pipeline that evaluates pull requests and assigns confidence scores to every potential issue. This code review confidence scoring system ensures reviewers receive only actionable, high-confidence findings by requiring each detected problem to meet a strict numerical threshold before surfacing.

The Confidence Scoring Pipeline

The scoring mechanism operates across four distinct phases, as implemented in plugins/code-review/commands/code-review.md and documented in plugins/code-review/README.md.

Parallel Agent Execution

The plugin launches four specialized agents simultaneously to analyze changes. According to the source code, these include two CLAUDE.md compliance agents, one Opus bug-detector, and one historical-context analyzer. Each agent independently produces a list of candidate issues, creating redundant coverage that reduces false positives.

Sub-Agent Validation

For every issue flagged by the bug-detector agents (agents 3 and 4), parallel sub-agents re-evaluate the finding against the PR title and description. Only issues that survive this secondary validation move forward to the scoring phase, eliminating speculative detections early.

Nx Confidence Scoring

Each validated issue receives individual attention from an Nx confidence scorer. This component analyzes the evidence—including code diffs, CLAUDE.md rule citations, and git-blame data—to calculate a numeric confidence value between 0 and 100. The scoring rubric documented in the README at lines 78-84 defines the scale:

• 0 – Not confident / false positive
• 25 – Somewhat confident
• 50 – Moderately confident
• 75 – Highly confident
• 100 – Absolutely certain

Threshold Filtering

After scoring, the plugin filters out any issue whose confidence falls below the configured threshold. The default threshold is 80, as specified in plugins/code-review/commands/code-review.md at lines 16-19 and documented in the README at lines 24-26. Only issues scoring 80 or higher appear in the final output.

Configuring the Confidence Threshold

Users can adjust the strictness of the filter by modifying the command definition file. To change the threshold, edit plugins/code-review/commands/code-review.md:

# In code-review.md (around line 16)

Filter out any issues with a score less than 80.

Replace 80 with any value between 0 and 100. Lower values surface more potential issues but increase noise, while higher values enforce stricter quality gates.

Running Code Reviews with Confidence Filtering

When you execute the review command, the confidence threshold automatically applies to all detected issues.

Run a review with the default 80-point threshold:

/code-review

Post high-confidence issues directly to the pull request:

/code-review --comment

Both commands only surface issues meeting the confidence threshold, ensuring the output contains exclusively high-signal feedback. If no issues meet the threshold, the plugin emits a "No issues found" message.

Summary

• Claude Code's code review plugin uses four parallel agents to detect potential issues in pull requests
• Each issue undergoes sub-agent validation before receiving a confidence score
• The Nx confidence scorer assigns values from 0-100 based on evidence quality, following the rubric in plugins/code-review/README.md lines 78-84
• A configurable threshold (default 80, defined in plugins/code-review/commands/code-review.md lines 16-19) filters out low-confidence results
• Only issues scoring 80 or higher appear in terminal output or PR comments

Frequently Asked Questions

What is the default confidence threshold in Claude Code?

The default confidence threshold is 80. Issues scoring below this value are automatically filtered out and never shown to users. You can modify this threshold by editing the filter value in plugins/code-review/commands/code-review.md at lines 16-19.

How are confidence scores calculated for detected issues?

An Nx confidence scorer evaluates each validated issue individually, examining evidence such as the code diff, specific CLAUDE.md rule citations, and git-blame data. It outputs a numeric score between 0 and 100 according to the rubric defined in plugins/code-review/README.md at lines 78-84.

Why does Claude Code use multiple agents for code review?

The plugin runs four specialized agents in parallel—two CLAUDE.md compliance checkers, one Opus bug-detector, and one historical-context analyzer—to create redundant coverage. This multi-agent approach reduces false positives and ensures comprehensive analysis before confidence scoring occurs.

Can I see all issues regardless of confidence score?

While the default behavior filters out issues below the threshold, you can lower the threshold to 0 in plugins/code-review/commands/code-review.md to surface all detected issues. However, this will include speculative findings and potential false positives that the confidence scoring system normally suppresses.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}