Code review

Code Review Workflow

Use this workflow to review git changes with evidence-backed findings and risk-aware verdicts.

Step 1: Determine Review Target

Identify what to review from the user's request.

User Input	Action
Commit hash provided	Use that commit
Commit range (start~end)	Use the range (start = earliest, end = latest)
"리뷰해줘" / "review" with no hash	Use staged changes (git diff --cached). If nothing staged, use HEAD
File paths specified	Scope review to those files within the target

Always confirm the target before proceeding:

git --no-pager show --stat <commit_hash>
# or for staged changes:
git --no-pager diff --cached --stat

Step 2: Retrieve the Diff

Get the full diff for the review target. Refer to git_operations.md for detailed commands and fallback handling (root commits, merge commits, range stabilization).

# Single commit
git --no-pager show <commit_hash>

# Staged changes
git --no-pager diff --cached

# Commit range
git --no-pager diff <start_hash>^..<end_hash>

# File-scoped
git --no-pager show <commit_hash> -- <file_path>

Step 3: Classify Changes and Filter Noise

Categorize each changed file and filter out noise before deep review. This prevents wasting review effort on non-reviewable content.

File Categories

Category	Examples
Source	.ts, .js, .py, .java, .go, .rs, .swift, .kt, .rb, .php, .c, .cpp
Test	*.test.*, *.spec.*, __tests__/*, test_*.*, *_test.*
Config	.json, .yaml, .toml, .env*, .*rc, config/*
Docs	.md, .rst, .txt, docs/*
Build/CI	Dockerfile*, Makefile, .github/workflows/*, *.gradle*
Deps	package.json, pyproject.toml, go.mod, Cargo.toml, requirements*.txt

Noise Filtering

Skip line-by-line review for these — check consistency/integrity only:

Noise Type	Pattern	Handling
Generated files	generated/, *.pb.*, *.generated.*	Inspect generation source instead
Lock files	package-lock.json, yarn.lock, go.sum, Cargo.lock	Verify manifest consistency only
Minified/bundled	*.min.js, *.bundle.js	Exclude from logic findings
Vendor/third-party	vendor/, third_party/	Evaluate integration impact only
Mechanical renames	Status R/C with no behavioral delta	Validate paths, then deprioritize
Format-only changes	Whitespace/indent-only diffs	Note as low-risk, skip deep analysis

Review Depth by Scale

Files Changed	Strategy
1–10	Full line-by-line review
11–30	Prioritize source/config, then tests/docs
30+	Separate noise first, sample low-risk files, full-review high-risk files

Always Full-Review (Regardless of Scale)

• Security/auth/authorization files
• Data schema, migration, persistence contracts
• Runtime entrypoints and routing bindings
• Critical production config files

Step 4: Analyze Reviewable Files

Read the full context of each reviewable file to understand the change, not just the diff. Focus on:

• What behavior changed and why
• Whether the change is correct and complete
• Edge cases, error handling, boundary conditions
• Consistency with surrounding code patterns

Step 5: Trace Impact

Identify who/what is affected by the changes. Refer to impact_detection.md for detailed techniques.

Key actions:

1. Find consumers — search for references to changed symbols, APIs, contracts
2. Check exposure — is the changed entity publicly consumed or internal-only?
3. Detect breaking changes — required inputs increased? Output contract changed? Symbols removed/renamed?
4. Check test coverage — do tests cover the changed behavior paths?

# Find direct callers
rg -F "symbolName(" .
# Find broader references
rg "symbolName" .

Breaking Change Signals

Change Type	Breaking?
New mandatory parameter/field	Yes
Removed/renamed output field	Yes
Accepted values narrowed	Yes
Endpoint signature changed	Yes
Additive optional extension	No (usually)

If a breaking change has normalized consumers > 0, classify at least as Critical candidate.

Step 6: Classify Findings

Assign each finding a severity and confidence level.

Severity Levels

Critical

Immediate high risk to security, data integrity, availability, or external contracts.

• Security exposure: injection, credential leakage, auth bypass
• Data integrity: corruption, irreversible mutation, destructive migration without rollback
• Availability: crash loops, deadlock, unbounded resource exhaustion
• Breaking contract: incompatible changes on consumed public interfaces
• Rollback gap: no rollback path for high-impact operational changes

Major

Incorrect behavior, reliability degradation, or high operational cost.

• Behavioral defects: wrong branch logic, boundary errors, invalid fallbacks
• Performance: unbounded processing, repeated expensive ops
• Error handling gaps: swallowed failures, incorrect retries
• Concurrency/race conditions
• Observability regression: lost diagnostic context
• Config semantics drift without compatibility handling

Minor

Maintainability, readability, or medium-term quality concerns.

• Excessive complexity, duplicated logic
• Ambiguous naming, unclear control flow
• Tight coupling, low cohesion

Nit

Low-impact polish and consistency.

• Style/formatting drift
• Naming clarity improvements
• Outdated comments

Critical vs Major Boundary

Factor	→ Critical	→ Major
Data risk	Loss/corruption likely, irreversible	Inconsistency possible, recoverable
Business logic	Core transaction/auth broken	Limited-path incorrect behavior
Rollback	No safe rollback path	Rollback exists and is practical
Observability	Incident detection critically degraded	Degraded but manageable

Confidence Tiers

Tier	Meaning
High	Strong direct evidence; severity can be acted on directly
Medium	Credible but partially indirect evidence; include verification note
Low	Weak or indirect evidence; avoid automatic escalation, request manual verification

Escalation Rule

A Major finding becomes a REQUEST_CHANGES candidate when ALL of:

• Impacts a critical domain (auth, payment, data integrity, availability)
• High user/service/data impact
• Confidence is Medium or High

Step 7: Determine Verdict

Combine findings into a final verdict.

Baseline Rules

Verdict	Condition
REQUEST_CHANGES	Any Critical finding
REQUEST_CHANGES	3+ Major findings
COMMENT	1–2 Major findings
COMMENT	5+ Minor findings
APPROVE	Only Minor/Nit or no findings

Risk-Weighted Adjustment (Internal Logic)

Use internally to validate the baseline verdict — do NOT output the score:

• Severity weights: Critical=10, Major=5, Minor=2, Nit=1
• Confidence factor: High=1.0, Medium=0.7, Low=0.4
• Critical domain bonus: +4 per finding in auth/payment/data integrity/availability

Score	Adjustment
≥ 12 with medium+ confidence	REQUEST_CHANGES
6–11	COMMENT (unless baseline already requests changes)
≤ 5 and no Major+ findings	APPROVE candidate

Confidence-Aware Handling

Evidence Shape	Handling
High confidence + high impact	Keep or escalate severity
Medium confidence + medium/high impact	Keep severity, add verification note
Low confidence	No automatic escalation; request manual verification

Step 8: Generate Report

Output the review using the format in output_format.md.

Key requirements:

• Group findings by severity (default) or by file if requested
• Each finding must include: file location, issue, evidence, impact, confidence tier, and suggestion
• State analysis limitations honestly (unverifiable areas, sampling scope)
• Include a Highlights section when patch quality deserves recognition
• End with a clear verdict and decision rationale

Tools

• Git CLI: Commit metadata, diffs, history, range context
• Search tools (rg): Consumer tracing, reference lookup, test coverage checks

Reference Files

File	When to Read
git_operations.md	Step 2 — for edge cases (root commits, merges, range fallbacks)
impact_detection.md	Step 5 — for consumer tracing and breaking change detection
output_format.md	Step 8 — for report structure and finding entry format