AI PR Reviewers: Diff-Scoped Context Windows That Catch Real Bugs

Full-file context makes AI PR bots verbose and blind: they nibble import order while missing an authz bypass in a callee two hops away. The unfair advantage is diff-scoped windows: changed symbols, call-graph neighbors, and contract tests only — then ask for nullability, authorization, and schema risks. Humans still own merge; the bot owns high-signal nits that would otherwise escape.

⚡ TL;DR: Build context from git diff + symbol extraction + 1–2 hop callers/callees + related *.test.ts / OpenAPI slices. Cap tokens hard. Prompt for ranked findings with file:line and severity. Drop style nits CI already covers. Wire secrets/IAM carefully (AI code review bots). Align with Cursor monorepo rules.

Build the context pack from the diff

// ci/ai-review/context.ts
import { execSync } from "child_process";

export function changedFiles(base: string): string[] {
  return execSync(`git diff --name-only ${base}...HEAD`, { encoding: "utf8" })
    .trim()
    .split("\n")
    .filter((f) => /\.(ts|tsx|js|py)$/.test(f));
}

export function unifiedDiff(base: string, file: string): string {
  // ✅ Prefer unified diff hunks over whole files
  return execSync(`git diff -U5 ${base}...HEAD -- ${file}`, { encoding: "utf8" });
}

// ❌ Dumping entire files "so the model has context"
const context = files.map((f) => readFileSync(f, "utf8")).join("\n\n");

Expand to call-graph neighbors (bounded)

export type SymbolHit = { file: string; name: string; kind: "fn" | "class" };

export function neighborContext(hits: SymbolHit[], maxHops = 1, tokenBudget = 6000): string {
  const parts: string[] = [];
  let used = 0;
  for (const h of hits) {
    const callees = lookupCallees(h).slice(0, 5); // ✅ hard cap
    const callers = lookupCallers(h).slice(0, 5);
    for (const n of [...callees, ...callers]) {
      const snippet = readSymbolSource(n);
      if (used + snippet.length > tokenBudget) return parts.join("\n\n");
      parts.push(`// neighbor ${n.file}::${n.name}\n${snippet}`);
      used += snippet.length;
    }
  }
  return parts.join("\n\n");
}

Include contract tests that mention the same symbols — that is where schema and authz regressions show up.

Prompt for bugs CI cannot see

You are a senior PR reviewer. Given DIFF hunks, NEIGHBORS, and TESTS:
- Find nullability, authz, tenancy, and schema/contract breaks
- Cite file:line and severity (blocker/major/nit)
- Ignore formatting, import order, and naming already enforced by eslint
- If unsure, say uncertain — do not invent APIs
Return JSON: { findings: [{ severity, file, line, title, why }] }

type Finding = {
  severity: "blocker" | "major" | "nit";
  file: string;
  line: number;
  title: string;
  why: string;
};

export function filterFindings(raw: Finding[]): Finding[] {
  // ✅ Drop nits; keep blockers/majors for human attention
  return raw.filter((f) => f.severity !== "nit" && f.line > 0);
}

Post comments without spamming

One summary comment + inline notes on blockers. Dedipe against previous bot runs on the same SHA.

# ✅ Idempotent review marker
gh api repos/$REPO/commits/$SHA/comments -f body="$(cat review.md)"

Security posture for tokens and least privilege: AI Code Review Bots: IAM, Secrets, and Least-Privilege Pipelines. Sandbox any tool execution: LLM Coding Agents on AWS.

What “good” looks like in metrics

Closing checklist

✅ Dos
– ✅ Context = diff hunks + bounded neighbors + contract tests
– ✅ Hard token budget; JSON findings with file:line
– ✅ Filter to blocker/major; suppress eslint-covered nits
– ✅ Least-privilege GitHub App tokens
– ✅ Track precision with human feedback

❌ Don’ts
– ❌ Don’t send entire unchanged files
– ❌ Don’t let the bot bikeshed formatting
– ❌ Don’t auto-approve merges from the model
– ❌ Don’t expand call graph unbounded
– ❌ Don’t paste secrets from CI into the model prompt

Related reading

• AI Code Review Bots: IAM, Secrets, and Least-Privilege Pipelines
• Cursor Rules for TypeScript Monorepos: Make AI Edits Stick
• LLM Coding Agents on AWS: Safe Tool Sandboxes with Lambda
• Amazon Bedrock Agents: Tool Use, Memory, and Production Guardrails

Last updated on September 11, 2026