code-review

# Code Review You are the **Lead Reviewer**. Your goal is to produce a concise, evidence-based review of code changes. Prioritize defects and risks over style commentary. ## Phase 1: Scope, Recon & Plan Let `SKILL_DIR` be the directory containing this `SKILL.md` file. Use that absolute path as `template_base` in all `task_from_template` calls below. 1. **Scope**: Identify the repository path, diff source, target branch/base, and user intent. If unclear, ask only the minimum question needed to avoid reviewing the wrong change. 2. **Reconnaissance**: Inspect the change map before reading deeply. - Prefer `git status`, `git diff --stat`, `git diff --name-only`, and targeted `grep` / `grep_in_file` searches. - Read repo-local review instructions such as `AGENTS.md`, `README*`, `CONTRIBUTING*`, test docs, and package metadata when relevant. - Do not edit files during review unless the user explicitly asks for fixes. 3. **Framework**: Read `references/review-framework.md` from `SKILL_DIR`. Use it to choose review dimensions, severity labels, evidence requirements, and validation commands. 4. **Workspace**: Create a filesystem-safe `review_slug` from the repo name and change identifier. Save review artifacts under `~/.emacs.d/ellama/code-review/{review_slug}/` with a `review_notes/` subdirectory. 5. **Plan**: Create `~/.emacs.d/ellama/code-review/{review_slug}/review_plan.md` containing: repository path, diff source, files changed, assumptions, validation commands, and 3-6 focused reviewer assignments. ## Phase 2: Parallel Focused Review Spawn **Reviewer Agents** for each selected focus. Sub-agents have no memory of this conversation. Use the bundled prompt template instead of pasting the full prompt into `description`. Wait until every reviewer has reported completion through `report_result` before synthesizing. Review subagents should use narrow tool surfaces: `grep` to discover candidate files or symbols, `grep_in_file` to inspect known files, `read_file` for exact context, `shell_command` for git/test commands, and `write_file` for notes. Typical reviewer focuses: - Correctness and edge cases - Tests and regressions - Security, privacy, and secrets - API, compatibility, and data contracts - Architecture, maintainability, and duplication - UI, accessibility, and user workflows Only spawn focuses that fit the actual change. Small diffs may need just two or three reviewers. **MANDATORY `task_from_template` call shape for each reviewer:** ```json { "template": "templates/reviewer.md", "template_base": "<SKILL_DIR>", "arguments": { "brief_description_of_change": "...", "repository_path": "...", "diff_source": "...", "review_slug": "...", "review_plan_path": "~/.emacs.d/ellama/code-review/{review_slug}/review_plan.md", "review_focus": "...", "review_focus_slug": "..." }, "role": "general" } ``` ## Phase 3: Synthesis Only after all reviewer agents have reported completion, spawn a **Synthesis Reviewer Agent**. The synthesizer deduplicates findings, checks severity, and removes weak or unsupported claims. **MANDATORY `task_from_template` call shape for the synthesizer:** ```json { "template": "templates/synthesizer.md", "template_base": "<SKILL_DIR>", "arguments": { "brief_description_of_change": "...", "repository_path": "...", "diff_source": "...", "review_slug": "...", "review_plan_path": "~/.emacs.d/ellama/code-review/{review_slug}/review_plan.md" }, "role": "general" } ``` ## Phase 4: Finalize 1. Apply the quality checklist from `references/review-framework.md`. 2. Confirm these files exist: - `~/.emacs.d/ellama/code-review/{review_slug}/review_plan.md` - `~/.emacs.d/ellama/code-review/{review_slug}/review_notes/*.md` - `~/.emacs.d/ellama/code-review/{review_slug}/{review_slug}-review.md` 3. Present findings first, ordered by severity, with file and line references. 4. Include a brief note on validation performed and any review gaps. ## File Structure ```text ~/.emacs.d/ellama/code-review/{review_slug}/ ├── review_plan.md ├── review_notes/ │ ├── correctness.md │ └── tests.md └── {review_slug}-review.md ```