Skip to content

Run an external-LLM code review

Run a code review on the modified files of a change using an external LLM that follows the global rules and the project's AGENTS.md. This is the fastest way to get a focused, convention-aware review of only the files touched by a change.

Prerequisites

  • A git repository with uncommitted or committed-but-unreviewed changes
  • Global rules present in your agent configuration (markdown files covering language, style, and documentation conventions)

Steps

Step 1 — Identify the modified files

git diff --name-only HEAD

If there are no uncommitted changes, compare against the base branch:

git diff --name-only main...HEAD

Step 2 — Gather the review context

Read the following to build the review context:

  1. The project's AGENTS.md (if present) — project-specific rules and conventions
  2. The global rules from your agent configuration — language rules, code style rules, documentation rules, commit rules
  3. The modified files from Step 1 — read each file in full

Step 3 — Run the external review

Send the modified files and the review context to an external LLM. The default model is swe; override via external_review_model in .zolletta-metaskill/settings.json or in the skill's front-matter.

The external reviewer should:

  • Follow the global rules and project AGENTS.md exactly
  • Review only the modified files (not the entire codebase)
  • Report findings with severity, file, line number, and suggested fix
  • Classify findings as actionable (require human judgment) or auto-fixable (formatter/linter could fix)

Step 4 — Collect the report

The external reviewer writes its findings to a markdown report in the timestamped report folder, following the standard report template with the grade at the top, tool results, auto-fixable issues, and findings grouped by severity.

Configuration

Setting Location Default Description
external_review_model .zolletta-metaskill/settings.json swe The external LLM model to use for the review

See also