Release / / with v2.3.4
code-review-graph v2.3.5
See how much review context the graph returns and verify its token estimate.
Before you upgrade to v2.3.5
1 to fixChange analysis can cap changed-function and transitive-test frontiers and return structured timeout errors using `CRG_TOOL_TIMEOUT`.
Read the evidence
Leiden community detection uses `CRG_LEIDEN_SEED=42` by default, making repeated grouping reproducible on the same graph; the environment variable selects another seed.
Read the evidence
Change benchmark configuration references from `nextjs.yaml` to `code-review-graph.yaml` and `eval --repo nextjs` to `eval --repo code-review-graph`; the old name described this repository, not Next.js.
What changed in v2.3.5
Brief change and update commands show a Token Savings panel with a category breakdown; `--verify` checks it against cl100k_base. Architecture output becomes compact by default.
Install code-review-graph 2.3.5 with pip; requires Python 3.10+. Optional integrations have separate dependencies.
New things you can do
Compare full and graph-filtered review context in a CLI panel and verify estimated tokens.
Read the evidence
Fixed and improved
The `embed` CLI command generates embeddings explicitly; richer embedding text and identifier-aware boosts improve retrieval.
Read the evidence
Windows local semantic search prewarms models to avoid a deadlock; missing local dependencies report provider unavailability.
Read the evidence
Cross-file callers remain visible even when local callers exist, user paths resolve to stored graph paths, and Rust test attributes produce test nodes.
Read the evidence
`build --skip-postprocess` and `update --skip-flows` no longer trigger an extra full post-processing pass.
Read the evidence
MCP, daemon and registry root validation consistently accepts SVN working copies.
Read the evidence
Seeded by @iluxav with Codex, from the release's public sources. Each change links to its evidence.