Skip to main content
Macroscope’s Correctness check reviews pull requests for bugs and posts findings on GitHub. In GitHub, you can reply to a finding with context, give feedback, or use Fix it for me to request a change. Choose a Detection mode to control coverage, validation, and cost. For custom checks with your own instructions, model, and tools, use Check Run Agents. Configure repository settings in Settings → Repos. Personal settings override repository preferences for the PRs you author. Defaults seeds newly linked repositories; it does not update existing ones.

​
Review triggers

In Settings → Repos, set Correctness to Always run, Manual invocation only, or Never run. Always run reviews PRs when they open or receive pushes. Manual invocation only reviews only when someone mentions @macroscope-app or adds an Always Review PR Label. Never run turns it off completely, and nothing overrides it. Onboarding turns Correctness on for the repositories you pick. Repositories linked later follow Defaults, which start off. The run-count and file-size limits are optional; blank or 0 clears them. Without a configured size limit, files of any size are reviewed.

​
PR labels

Two labels naming different modes prevent reviews from running. Removing a mode label does not itself trigger a review. Label-triggered reviews count as automatic: correctness spend caps and Max Automatic Runs per PR still apply. See Check Run Agent trigger rules for personal overrides and per-agent filters.

​
Minimum severity to comment

Set Minimum Severity to Comment to Low, Medium, High, Critical, or Off (default). Off posts all findings. Other values withhold findings below the threshold and findings without recorded severity. The PR author’s personal threshold overrides the repository threshold, including in CLI reviews. This affects correctness comments only. Withheld findings remain in the check’s Filtered Issues Details, tagged Below severity threshold, and remain available to Approvability. Minimum Blocking Severity is separate: it controls which outstanding correctness comments block automatic approval.

​
Issue severity levels

​
Manually triggering a review

A manual request bypasses automatic run-count and correctness spend caps; an always-review or detection mode label does not. A manual request does not bypass the workspace balance, monthly spend limit, ignore rules, or file-size rules. See spend controls.
Unless a Detection Mode is explicitly mentioned (or the author has a personal override configured), the review will use the repository’s configured Detection Mode.

​
Ignore rules and custom instructions

  • Ignore rules: exclude generated files, dependencies, or other paths; opt into reviewing tests with ignoreTests: false.
  • Custom instructions: explain intentional patterns and conventions so Correctness avoids incorrect comments, or wait for CI before reviewing.

​
Results

Correctness reports neutral when it finds issues. Review details list excluded files and why they were skipped. New pushes re-evaluate findings; resolved comments link to the commit that addressed them. The app’s Code Review activity log filters by source, including Correctness and Check Run Agents. Filters persist in the URL and apply to CSV exports. Feedback metrics can also be filtered by review type. Hover a Check Run Agent’s name on one of its comments to see the config commit the comment was generated from, and the commit that last changed the agent’s file as of that config commit.

​
Language support

Correctness reviews every language, plus configuration, documentation, and scripts, subject to file exclusions. It explores the surrounding code in your repository and can use web search for current library and API documentation.