code-review

name	code-review
description	Review a changeset using a repo's Copilot review instructions. Use when asked to review a PR, branch, diff, or set of local changes.

Step 1: Gather the applicable review prompts

Run the script with a start identifier and an optional stop identifier:

pkgx uv run <skilldir>/scripts/gather_prompts.py <start> [stop] --repo <path>

• <start> — required git identifier (commit, branch, or tag) to diff from.
• [stop] — optional ending identifier. If omitted, the diff runs from <start> to the working tree and also includes staged, unstaged, and untracked files — i.e. everything dirty in the working dir.
• --repo <path> — repository to inspect (default: current directory).

The script prints, one per line, the path of every instruction file that applies to the changeset: each .github/instructions/*.instructions.md whose applyTo globs match at least one changed file, plus the repository-wide .github/copilot-instructions.md. Files without an applyTo, or with excludeAgent: code-review, are skipped — matching Copilot's own selection rules.

Step 2: Spawn parallel review agents

Spawn one review agent per prompt file returned in Step 1, all in parallel, using the default subagent type. For each agent:

• Pass the prompt file as a context file (do not paste its contents into the prompt).
• Tell the agent the repo path and the diff range (<start> and optional [stop]).
• The sharedPrompt should ask subagents to review the actual diff (git diff <start> [stop]) only from the perspective of its single prompt file, reporting concrete, actionable findings with file/line references, and not restating the guidelines.

Step 3: Consolidate

Present the findings to the user, grouped or deduplicated as appropriate.