Onboard a monorepo to AI code review

A PR in a monorepo often touches 3-5 packages: API contract change in one package, consumer updates in others, test fixes in a fourth. Diff-only AI review sees each change in isolation, can't reason about the API contract's downstream consumers.

Result: the review approves changes that look fine in isolation but break callers in other packages. Bugs ship. The team learns to ignore the AI review.

Indexed monorepo review fixes this: the indexer builds a cross-package symbol graph at install time. PR review pulls the changed symbols + their downstream callers + relevant tests from the index. Agents see the FULL impact, not just the diff.

(1) Install the LGTM GitHub App on the monorepo. (2) First-index runs automatically — for typical monorepos (1k-10k files), 1-3 minutes; for large monorepos (100k+ files), 5-15 minutes. Progress streams to the dashboard via Socket.IO.

(3) Configuration: by default, every file gets indexed. To exclude paths (vendor/, third-party/, generated/, build/), create `.lgtmignore` at the monorepo root using gitignore syntax. Common exclusions: `node_modules/`, `dist/`, `*.pb.go`, `generated/`.

(4) Per-package overrides: Pro plan. Each package can have its own review configuration — different model, different focus areas, different verdict thresholds. Marketing-site package on Haiku 4.5 for cheap reviews; payments package on Opus 4.7 for thorough reviews.

A typical monorepo PR might touch 50 files across 5 packages — that's 5,000-50,000 lines of context if you tried to send it all to an LLM. Frontier models can handle that context, but the per-PR cost would balloon.

LGTM's approach: PageRank-ranked context retrieval. The diff's 50 changed files get embedded as graph seeds; personalized PageRank ranks the rest of the symbol graph by structural relevance. The top 15-30 symbols across all packages get attached to the prompt, NOT all 50,000 lines.

Net result: a 50-file monorepo PR review costs about the same in tokens as a 5-file single-package PR (the context budget is bounded), but the agents see the cross-package dependencies that matter.

Different packages have different needs. The auth-service package might need strict Security gating; the marketing-site package might only need Readability.

Configure per-package: Settings → Repos → Per-Path Config. Patterns are glob-based (`packages/auth/**`, `apps/marketing/**`). Each pattern can specify: enabled agents, model override, verdict threshold, ignore patterns.

Inheritance: package config inherits from repo defaults; repo config inherits from org defaults. Effective config is computed on every review by walking the file path and matching patterns longest-first.

Naive math: 50 PRs/week × 10 engineers × 6 agents × ~3,000 tokens per agent × $0.005/1k tokens = ~$45/week ≈ $180/month in pure tokens. That's the worst-case for a busy monorepo on gpt-5.4.

Optimized math: per-package model tuning (mini/haiku/nano on low-stakes packages, flagship only on critical packages) drops average per-PR cost by 60-80%. Realistic: $30-$60/month in tokens for a busy 10-engineer monorepo.

LGTM's flat subscription (₹399/mo Pro) is separate from token cost. Combined: ~$30-$80/month total for unlimited reviews on a busy monorepo. No per-seat tax, no per-repo tax.