code-review

Reviews multi-file code changes using graph-based blast-radius analysis and confidence-scored findings.

Updated Mar 22, 2026
One-click install
npx skills add https://github.com/diazMelgarejo/orama-system --skill code-review-diazmelgarejo
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: code-review
Source: https://github.com/diazMelgarejo/orama-system/tree/main/bin/orama-system/skills/code-review
Command: npx skills add https://github.com/diazMelgarejo/orama-system --skill code-review-diazmelgarejo

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires jq, code-review-graph, and includes scripts (resource) and references (resource) components.

What problem does it solve? Reviewing multi-file changes and pull requests by reading files inline wastes tokens and misses cross-module impact. This Skill enforces a graph-first review chain that maps blast radius before reading code, so reviewers only open confirmed-relevant files and report verified findings. ## Core Features & Use Cases - Graph-first review chain: Runs code-review-graph (detect_changes, impact radius, review context), then gbrain symbol resolution, then scoped file reads — never whole-repo scans. - Delta and PR modes: Single-pass review for local diffs; five-lens parallel fan-out (guidelines, shallow bugs, git history, prior PRs, in-file guidance) for full PR reviews. - Confidence-gated reporting: Every finding is scored 0-100; only issues at 80 or above reach the report, with Critical/Important severity tiers and file:line citations. - Use Case: Before merging a branch that touches a shared API contract, run the PR mode to map the blast radius, fan out lens workers across orchestration backends, and receive a merged, deduplicated verdict of Yes, No, or With fixes. ## Quick Start Ask the agent to review your uncommitted changes or a specific pull request using the code-review skill, for example: review my uncommitted delta for bugs before I commit.

Frequently Asked Questions about code-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I review a pull request with code-review-graph?

Start with detect_changes_tool and get_impact_radius_tool to build the assigned file list, then fan out the five review lenses (guidelines, shallow bugs, git history, prior PRs, in-file guidance) via OmniRoute, ai-cli, or Task subagents. Merge results and drop anything below confidence 80.

What is the correct tool order for reviewing code changes?

The non-negotiable chain is code-review-graph first for blast-radius and change detection, then gbrain for symbol definitions and past decisions, then Read only on graph-confirmed files. Never run whole-repo reads or broad Grep sweeps before the graph step.

Does this code review skill work with Cursor and Codex?

Yes, the same graph-first chain applies across Claude Code, Cursor, Codex, Gemini, OpenClaw, and Hermes. Only the invocation differs: Cursor and Claude Code use MCP tools directly, while Codex calls them via codex mcp call or ai-cli workers.

Why is the code-review-graph MCP server disconnected?

The first uvx cold start downloads tree-sitter packages and can exceed the MCP handshake window. Pre-warm the cache with uvx code-review-graph --help, then reconnect via /mcp. If it stays down, fall back to gbrain search, which shares the same bge-m3 vector space.

When should I use Delta mode instead of PR mode?

Use Delta mode for uncommitted changes, small diffs under roughly ten files, and pre-commit checks with a single reviewer pass. Use PR mode for branch-versus-main reviews, large diffs, or explicit multi-lens requests that justify parallel fan-out.

Why did my review report drop some findings?

Findings below confidence 80 are filtered out by design. The rubric drops pre-existing issues, linter-catchable errors, pedantic style nitpicks, and speculative bugs that lack a concrete file:line, trigger input, and explanation of why existing guards do not catch them.