Pr Review

PR Review Skill

Reviews code changes with focus on quality, security, and consistency.

Default Assumption: Public Repository

Unless explicitly stated otherwise, assume the repository is publicly available. This means:

• Any secret, credential, or API key pushed is considered compromised
• Internal URLs, IPs, hostnames should not be exposed
• Comments with sensitive internal context should be flagged
• Error messages should not leak internal architecture
• Be extra cautious with .env files, config files, CI/CD configs

Phase 1: Determine Scope

STOP. Use AskUserQuestion before anything else.

Ask user to choose review scope:

• Staged files only
• Unstaged changes (working directory)
• All uncommitted (staged + unstaged)
• Current branch vs main (PR-style)
• Specific commit or range
• Other (specify)

Do NOT run any git commands or tools until user responds.

After selection, get the diff:

• Staged: git diff --cached
• Unstaged: git diff
• All uncommitted: git diff HEAD
• Branch vs main: git diff main...HEAD
• Commit: git show <hash>
• Range: git diff <from>..<to>

Also get changed files list: git diff --name-only <appropriate args>

Phase 2: Understand the Problem

STOP. Use AskUserQuestion to confirm before proceeding.

Infer intent from:

1. Branch name: git branch --show-current
2. Commit messages: git log main..HEAD --oneline (or relevant range)

Then use AskUserQuestion to confirm:

"Based on branch feature/xyz and commits, this PR appears to [inferred description]. Is this correct?"

• Yes, proceed
• No, let me explain

Do NOT proceed until user confirms.

Phase 3: Auto-Detect Stack

Check for presence of:

• package.json / yarn.lock → Node.js
• requirements.txt / pyproject.toml → Python
• go.mod → Go
• Cargo.toml → Rust
• Dockerfile → Docker
• *.tf → Terraform
• *.yaml in k8s patterns → Kubernetes
• .github/workflows/ → GitHub Actions

Note detected stack for context-aware analysis.

Phase 4: Triage (large diffs only)

If git diff --stat shows more than ~20 files or a few thousand lines, do not read every file. Reading a huge diff end-to-end produces a shallow review of everything instead of a deep review of what matters.

Instead:

1. Group the changed paths by area, service, or module.
2. Flag the high-risk areas first: deploy manifests, Helm/Kustomize, Terraform/IaC, CI/CD pipelines, Dockerfiles, config/env/secrets, migrations, feature flags, networking, and anything touching runtime startup or health checks.
3. Give a tight per-area summary — line counts and what each area appears to do.
4. Use AskUserQuestion to ask which areas to review deeply. Wait for the answer.

Then scope the remaining phases to the selected paths: git diff main...HEAD -- <path>.

Skip this phase entirely for small diffs — the gate costs a round trip and buys nothing.

Phase 5: Run Scanners

Execute relevant scanners (skip silently if not installed):

Always run:

Stack-specific:

Phase 6: Code Review

Analyze the diff for all categories. Be pragmatic—flag likely issues, skip obvious false positives.

When the diff touches infrastructure, deploy config, or anything that runs in production, lead the report with 6.8 — operational risk outranks style.

6.1 Code Quality

• Best practices for detected stack
• Readability and maintainability
• Error handling appropriateness
• Test coverage (if tests exist)
• Idiomatic patterns
• Type safety issues

6.2 Codebase Consistency

• Match existing patterns in the repo
• Naming conventions alignment
• File organization consistency
• Don't introduce a 10th way of doing something

6.3 Security

Manual checks:

• Hardcoded secrets, API keys, passwords, connection strings
• SQL injection, XSS, command injection vectors
• Path traversal risks
• Auth/authz bypasses
• Insecure defaults (http vs https, weak crypto)
• Sensitive data in logs/errors/URLs
• Container: running as root, privileged mode, unverified base images

6.4 Bug Detection

• Logic errors, off-by-one
• Null/undefined handling
• Race conditions
• Resource leaks (unclosed handles, connections)
• Breaking changes to existing APIs

6.5 Dependencies

• Known vulnerable package versions
• Outdated dependencies with security patches
• Unpinned versions
• Suspicious or typosquatted package names

6.6 Performance

• N+1 query patterns
• Sync operations in async contexts
• Unbounded loops/recursion
• Memory leaks
• Missing pagination
• Blocking I/O in hot paths

6.7 Deprecations & Drift

• Deprecated APIs, functions, patterns
• Breaking changes in dependencies
• Hardcoded values that should be variables
• Environment-specific configs in shared code
• Configuration diverging from IaC patterns

6.8 Operational Risk (SRE lens)

Ask what this change does to a running production system, not just whether the code is correct.

• Blast radius & rollback: what breaks in prod if this is wrong, how hard is it to revert, is it backward-compatible with the currently-deployed version?
• Deploy & runtime risk: config changes, resource limits, startup/liveness/readiness probes, secrets handling.
• Failure modes: retries, timeouts, error handling, partial-failure behavior, idempotency.
• Observability: are the new code paths logged, metered, traced? Can this be debugged at 3am by someone who did not write it?
• Migrations & data: ordering, reversibility, downtime, dual-write/dual-read windows.
• Scaling & cost: hot paths, unbounded loops or memory, new external calls, added per-request latency.

When reviewing an unfamiliar codebase, explain the domain context behind a finding rather than assuming it — and say explicitly when something needs the original author or a domain expert rather than guessing.

Phase 7: Report

Output a succinct markdown report:

## PR Review: [brief title]

**Problem:** [1-2 sentences on what this PR solves]

**Scope:** [staged/branch/commits reviewed]

**Stack:** [detected tech stack]

### Scanner Results
| Tool | Result |
|------|--------|
| gitleaks | [clean/N findings] |
| ... | ... |

### Findings

#### CRITICAL
- `file:line` - [issue with brief context]

#### HIGH
- `file:line` - [issue]

#### MEDIUM
- `file:line` - [issue]

#### LOW
- `file:line` - [issue]

### Summary
- Critical: X | High: X | Medium: X | Low: X

### Review Score: X/20
[One sentence justification]

### Action Required
| Priority | Item |
|----------|------|
| blocker | ... |
| should fix | ... |
| consider | ... |

Rating Scale

Style Guidelines

Keep findings concise but contextual:

• Bad: "should use https here"

• Good: "http exposes data in transit, use https"

• Bad: "fix this null check"

• Good: "user.email accessed without null check - crashes if user not found"

Don't write a 50-page report. Focus on what matters.