Structural PR Analysis That Makes AI Review Better

The Problem: AI Code Review Doesn't Scale

What CKB Computes (0 Tokens)

These questions have deterministic answers that don't need an LLM. CKB answers all of them using SCIP indexes, tree-sitter, and git history.

Total: CKB answers in ~1K tokens what would cost an LLM ~790K tokens to compute from source. Two of these questions (churn history, co-change patterns) are impossible for an LLM without git access.

How It Works

What You Get

CKB Review: WARN  ·  133 files  ·  19,200 lines
========================================================

  Changes 133 files across 32 modules (go). 8 new bug
  pattern(s); 1 co-changed file missing. Consider splitting
  into 32 smaller PRs.

Checks:
  ⚠ bug-patterns     8 new bug pattern(s) (31 pre-existing filtered)
  ⚠ coupling         1 co-changed file missing
  ⚠ risk             Risk score: 1.00 (high)
  ⚠ split            32 independent clusters
  ○ test-gaps        16 untested functions (top 10 shown)
  ○ hotspots         50 hotspot files (top 10 shown)
  ✓ secrets · breaking · dead-code · health · tests · complexity
    · format-consistency · comment-drift

Top Findings:
  ⚠ review.go:267       err shadowed (outer error lost)
  ⚠ setup.go:215        err shadowed (non-fatal)
  ⚠ diff.go             +11 cyclomatic in GetCommitRangeDiff()
  ⚠ pr.go               +13 cyclomatic in SummarizePR()
  ... and 27 more findings

Code Health: 0 degraded · 8 new (avg 82)
Estimated Review: ~3 hours (133 files, 19K lines)
Reviewers: talantyyr (9%) · lisa.welsch (9%)

20 checks, 5.2 seconds, 0 tokens, 0 API calls. Compact mode (~1K tokens) gives AI assistants exactly what they need for triage.

20 Quality Checks

All checks run concurrently. Tree-sitter calls are serialized via a mutex; git subprocess work overlaps with static analysis. 8 checks work without SCIP indexing; add ckb index for the full 20.

Benchmark Data

All numbers from real measurements on a production PR (133 files, 19,200 lines, 37 commits).

Flexible Invocation

# Full PR review
ckb review --base=main

# Staged changes only (pre-commit)
ckb review --staged

# Scope to a directory
ckb review internal/auth/

# Scope to a symbol
ckb review --scope=AuthMiddleware

# Specific checks only
ckb review --checks=breaking,secrets,bug-patterns

# CI mode (exit codes: 0=pass, 1=fail, 2=warn)
ckb review --ci --format=sarif

# Post as PR comment
ckb review --base=main --post=123

# Via MCP (any AI tool)
reviewPR(baseBranch: "main", compact: true)

7 Output Formats

CI Integration

- uses: actions/checkout@v4
  with:
    fetch-depth: 0

- name: CKB Review
  run: |
    npx @tastehub/ckb review \
      --base=${{ github.event.pull_request.base.ref }} \
      --ci --format=sarif > review.sarif

- uses: github/codeql-action/upload-sarif@v3
  with:
    sarif_file: review.sarif

Performance

Parallel check execution, batched git operations, cached hotspot scores, and mutex-based tree-sitter scheduling. The bottleneck is tree-sitter complexity analysis (~1.8s) and coupling analysis (~1.8s).

What It Does — and Doesn't

CKB Review is a structural pre-filter, not a replacement for code review. It tells your AI reviewer where the problems are. Finding what the problems are still requires reading the code.