tirth8205/code-review-graph
7 skills5.1k installs128k starsGitHub
Install
npx skills add https://github.com/tirth8205/code-review-graphSkills in this repo
1Build GraphBuild Graph is an agent skill for the code-review-graph MCP stack that creates or incrementally updates a persistent repository knowledge graph stored as SQLite under .code-review-graph/graph.db. Solo builders and small teams install it when they want review agents to reason over nodes and edges—files, symbols, and relationships—rather than re-scanning the tree each session. The documented flow checks status with list_graph_stats_tool, runs build_or_update_graph_tool (full rebuild on first run, incremental afterward), then re-lists stats to report files parsed, nodes and edges, languages detected, and errors. Hooks keep the graph aligned with edits and commits, so manual invocation is mainly for first-time setup, major refactors, branch switches, or when the graph feels out of sync. Binary, generated, and ignored paths are skipped per .code-review-graphignore, which keeps graphs lean for real review workloads across a wide multi-language codebase.817installs2Review DeltaReview Delta is an agent skill for focused, token-efficient code review of only what changed since the last commit. It assumes a code-review graph toolchain: refresh the graph incrementally, pull review context with changed snippets and impacted nodes/files, then review correctness, style, and blast radius while checking whether dependents and tests need updates. Solo and indie builders use it when diffs are small but risk is high—signature changes, inheritance tweaks, or edits in heavily depended-on modules—without paying for a full-repo review pass. The skill encodes a repeatable sequence: ensure graph currency, analyze impacted_nodes and impacted_files, review each changed file against guidance, and verify test coverage gaps. It pairs naturally with git-based ship workflows in Claude Code, Cursor, or Codex where the agent can invoke graph MCP tools alongside normal file reading.774installs3Review Prreview-pr is an agent skill that performs comprehensive code review on a pull request or branch diff using a codebase knowledge graph instead of reading every file ad hoc. Solo builders and small teams use it when they want merge decisions backed by structural context: which functions changed, who calls them, and whether tests exist. The skill walks through updating the graph from a main base, pulling full PR review context, analyzing impact radius across the diff, and deep-diving significant files with targeted graph queries. It emphasizes token efficiency by loading the optimized review-pr doc section first and avoiding full-file dumps unless asked. It fits Claude Code-style agents that expose build_or_update_graph_tool, get_review_context_tool, get_impact_radius_tool, and query_graph_tool. Output is a structured review with blast-radius analysis rather than informal chat feedback.768installs4Explore CodebaseExplore Codebase is an agent skill that standardizes how you navigate a repository through the code-review-graph MCP instead of blind grep or huge file dumps. Solo builders attach it when an agent must map architecture, modules, execution flows, and symbol relationships before a meaningful code review or targeted refactor. The procedure starts broad with list_graph_stats and get_architecture_overview, narrows via list_communities and semantic_search_nodes, then traces behavior with query_graph and list_flows. Token efficiency is a hard editorial constraint: always begin with get_minimal_context for the task, keep detail_level at minimal unless insufficient, and aim to finish orientation in at most five tool calls and eight hundred output tokens. It complements human review by giving the agent a graph-native mental model of callers, callees, imports, and large functions.722installs5Refactor SafelyRefactor Safely is a multi-phase agent skill for solo builders who maintain real codebases and want graph-backed refactoring instead of blind search-and-replace. It walks through refactor_tool in suggest, dead_code, and rename modes, then apply_refactor_tool using a refactor_id, and verifies with detect_changes. Safety checks emphasize previewing rename edit lists, checking impact radius and affected flows before large edits, and using find_large_functions for decomposition. Token efficiency rules require get_minimal_context first and minimal detail unless escalation is needed. It fits Prism’s Ship and Build shelves because refactoring spans implementation and pre-release review, especially when agents risk spanning too many files in one session.700installs6Review ChangesReview Changes is an agent skill that walks a solo builder through a knowledge-graph code review instead of skimming a flat diff. It starts with change detection and risk scoring, maps affected execution flows, checks whether high-risk functions have tests via graph queries, and measures blast radius before suggesting targeted tests where coverage is missing. Output is organized by risk level with what changed, why it matters, coverage status, improvements, and an overall merge call. The skill encodes strict token efficiency: always prime context with get_minimal_context, keep detail_level minimal unless you must escalate, and aim to finish in five tool calls and under eight hundred output tokens. It fits indie teams shipping backend or API code who already use the code-review-graph tooling and want review steps that mirror impact analysis, not generic lint commentary.685installs7Debug Issuedebug issue is an agent skill that teaches a repeatable debugging loop on top of a code knowledge graph. Solo builders use it when a bug report is vague, a regression landed after a merge, or a change might ripple farther than one file. The procedure starts by pulling minimal context for the stated task, then semantically locating related nodes, tracing callers and callees, opening execution flows through suspected neighborhoods, checking detect_changes for fresh diffs, and measuring get_impact_radius on hot files. The skill encodes token discipline—minimal detail by default, at most five graph calls, and sub‑800 token answers—so agents do not drown in graph dumps during incident chat. It is best when your repo is already indexed for graph queries; it is not a substitute for writing reproduction tests. Use it in Operate when production misbehaves and in Ship when validating whether a risky PR could explain new failures.680installs