Local-first code intelligence graph for MCP and CLI. Builds a persistent map of your codebase so AI coding tools read only what matters, with benchmarked context reductions on reviews and large-repo workflows.
| Rating | Summary | |
|---|---|---|
| Quality | solid (19/24) | Actively maintained, well-benchmarked Python tool with massive early adoption and clean MCP/CLI surface |
| PAI Relevance | integrate (0.50) | Fills a genuine gap in code-aware context management; callable as a subprocess MCP server from PAI skills |
19/24 — actively-maintained / adequately-documented / solid
Failed:
Passed:
github.com/tirth8205/code-review-graph/actions/workflows/ci.ymlFailed:
Passed:
pip install code-review-graph and three follow-on commandscode-review-graph fixes that."docs/REPRODUCING.mdFailed:
Passed:
.github/workflows/ci.yml; README describes an automated evaluation runner with reproduction instructions in docs/REPRODUCING.md| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 1 | The blast-radius algorithm (trace callers/dependents/tests from a changed node) and Leiden-community-based context segmentation are both worth studying as design patterns for PAI's Knowledge skill, which currently handles general knowledge graphs without code-structure awareness. |
| Integration Readiness | 1 | Python implementation, but exposes a clean CLI (code-review-graph) and MCP server; PAI could wrap it as a subprocess skill or invoke it as an MCP tool from any agent — same integration path as other Python CLIs in PAI's toolchain. |
| Overlap Risk | 1 | Partial conceptual overlap with PAI's Knowledge skill (persistent graph querying), but code-review-graph is narrowly scoped to code-structure graphs, blast-radius analysis, and AST-level indexing — none of which the Knowledge skill covers. |
| Gap Fill | 1 | PAI has no code-aware context reduction capability; when PAI agents perform development tasks they lack blast-radius analysis to minimise file reads. This addresses a functional area with limited existing PAI coverage. |
Composite: 0.50
Aphoria (TypeScript, Bun, RiTa — Stage 3A complete, further stages in progress): Run pip install code-review-graph && crg index at the Aphoria project root, then register the MCP server in Claude Code's ~/.claude/mcp.json. When starting each new stage, Claude Code resolves the blast radius around the module being extended rather than scanning the full RiTa integration and generator tree — context per session shrinks to the structural neighbourhood of your entry point, updating incrementally as files change without any per-session configuration.
Capture-to-Knowledge Pipeline's clean-room Haiku validation step: Insert crg blast-radius <target-file> as a pre-step in the validation pipeline that outputs a minimal affected-file manifest; pass that manifest to the Haiku prompt builder instead of the full repo snapshot. Token cost per validation run then scales with the size of the change rather than the size of the codebase, and the blast-radius output becomes a structured audit record of exactly which files were in scope — directly usable in the manual validation phase to confirm scope coverage.
Conservancy (SvelteKit, Cloudflare Pages) ongoing editorial pipeline work: Run crg index once at the project root and connect the MCP server to Cursor or Claude Code. Any session adding routes or extending pipeline components queries the graph for structurally adjacent files rather than re-reading all layouts and component files — replaces implicit "read whatever looks relevant" context gathering with a reproducible structural query, keeping AI behaviour consistent as the codebase grows beyond the current mature editorial state.
Category: Developer Tools & CLIs
In this category: casey--just; tirth8205--code-review-graph is the first entry in the code-intelligence graph sub-niche
Standing: Complements casey--just (task-runner automation) by solving a different problem — intelligent context scoping for AI-assisted code review — with no functional overlap between the two entries.
Density: 9/10 — Available: repo metadata (stars, forks, dates, license, topics), full README first 8KB including diagrams index and benchmark details, landscape context with rolling summary and prior appraisals, release data. Missing: dependency manifest (pyproject.toml not fetched), full README beyond 8KB (some sections likely truncated), contributor count.
The prior rolling-summary entry for this repo scored poor (4.6/24), likely assessed under minimal data before the full README was available. With evidence_density 9/10, the score revises substantially upward to 19/24 (solid). The benchmark methodology (fixed upstream SHAs, deterministic Leiden seed, CPU-reproducible embeddings) is unusually rigorous for a developer tool and is itself a harvest signal for how PAI's Evals skill could approach infrastructure-level reproducibility.