code-review-blast-radius

name: code-review-blast-radius description: >- Build a blast-radius code graph for AI-assisted code review. Maps call chains, dependency propagation, and community clusters from a code diff to prioritize which files ACTUALLY matter for review (100% recall, token-budget capped). Derived from tirth8205/code-review-graph (MIT, 8.4/10 port score). Core algorithms: tree-sitter AST → call graph → Leiden community detection → betweenness centrality → blast-radius ranked output. 38×–528× token reduction vs naive all-files review. Use before any non-trivial PR review or merge. status: stable lane: [code-review, analysis, ci] type: analysis-workflow trigger: >- "code review graph", "blast radius", "what does this change affect", "which files matter for this PR", "call graph review", "impact analysis", "review this diff with graph" source-ingest: .planning/ingest/2026-06-06-article-github-com-tirth8205-code-review-graph.md port-score: 8.4/10 argument-hint: --diff [--budget-tokens 8000]

code-review-blast-radius

AI-assisted code review using a blast-radius call graph. Prioritize the files that propagate changes furthest; ignore files that are purely local.

Derived from tirth8205/code-review-graph (MIT). Clean-room reimplementation of the Leiden + betweenness-centrality blast-radius algorithm in our stack.

Core insight (from RE)

Naive code review sends ALL changed files to the LLM — token-expensive and context-unfocused. The blast-radius approach:

1. Parse changed files → call graph (tree-sitter, per language)
2. Run Leiden community detection → find module clusters
3. Score each changed function by betweenness centrality in the community graph
4. Cap output by token budget — send ONLY the highest-centrality files first
5. Result: 38×–528× token reduction; 100% recall on high-impact changes

Workflow

1. Build the code graph (once per session/PR)

xp code-graph build --diff HEAD~1..HEAD   # or full repo: --path .
# Outputs: .planning/.code-graph/graph.db (SQLite, tree-sitter-derived)
# Languages: Rust, TypeScript, JavaScript, Python (via tree-sitter grammars)

2. Get blast radius for the diff

xp code-graph blast-radius --diff HEAD~1..HEAD --budget-tokens 8000
# Outputs ranked file list: (file, centrality_score, community_id, n_callers, n_callees)
# Example: crates/xp-recall/src/retrieval.rs (centrality=0.89, community=core-retrieval, callers=23)

3. Review in order

Review files in centrality rank order:

• Highest centrality first — these are the load-bearing changes
• Skip or skim files below centrality threshold (configurable, default 0.15)
• If a community cluster appears in multiple changed files — review entire cluster

4. MCP exposure (for Kael/agents)

The code graph exposes 30 MCP tools (mirroring code-review-graph's interface):

• code_graph_query(function_name) — get call chain for a function
• code_graph_community(file) — which community cluster does this file belong to
• code_graph_blast_radius(diff) — full blast-radius ranked output
• code_graph_callers(function) — who calls this function
• code_graph_callees(function) — what does this function call

5. Post-review: update graph

xp code-graph update --diff HEAD~1..HEAD  # incremental update after merge

Token budget control

The blast-radius output is token-budget-capped. Configure per context:

• --budget-tokens 4000 — tight (fast review, lowest cost)
• --budget-tokens 8000 — standard (recommended)
• --budget-tokens 20000 — deep review (full clusters exposed)

Leiden community detection

Leiden (improvement over Louvain) partitions the call graph into communities. A community = a set of functions that call each other more than they call outside. A changed function in a high-betweenness position bridges communities — changes there propagate everywhere. These are the highest-priority review targets.

Language support (tree-sitter grammars)

Anti-patterns

• Do NOT skip the graph build and just send all diff files to the LLM — this is the expensive baseline this skill replaces
• Do NOT set --budget-tokens above 20000 without a good reason — defeats the token-reduction purpose
• Do NOT re-build the full graph on every review — use xp code-graph update for incremental diffs after the initial build

Relationship to existing tools

Provenance

• Source: tirth8205/code-review-graph (MIT) — .planning/ingest/2026-06-06-article-github-com-tirth8205-code-review-graph.md
• Port score: 8.4/10
• Key algorithms: tree-sitter AST → SQLite call graph → Leiden community → betweenness centrality → blast-radius ranking
• Token reduction validated: 38×–528× in upstream benchmarks
• Clean-room: algorithm described from RE spec and README; no Python source copied
• Built: 2026-06-07