Skill: Change Re-Audit (Side Effects, Regression, Edge Cases)

Type: Execution

Purpose

Perform a rigorous re-audit of the current changes. Identify:

• Side effects introduced by the modification
• Hidden regression risks
• Unhandled edge cases
• Operational, performance, and security implications

The output must provide evidence-based findings and concrete mitigation steps.

When to Use

• After code changes are complete and before merging to a shared branch
• After complex refactoring that touches multiple modules
• Before deploying a hotfix to production
• When a previous review missed edge cases or side effects

When NOT to Use

• During initial implementation (code is still being written)
• Documentation-only or comment-only changes
• Auto-generated code changes (e.g., lock files, migration snapshots)
• Changes already fully audited through another review skill

Inputs Required

Do not run this skill without:

• ] Full diff of current working changes
• ] List of modified files
• ] Affected entry points (API handlers, jobs, CLI, consumers, etc.)
• ] Existing related tests (unit/integration/e2e)

Optional but recommended:

• ] Architectural context (system diagram, dependency map)
• ] Deployment target information (staging, production, canary)

Without asking the user (unless unavailable), gather these inputs directly from the repository.

Output Format

1. Change Units
2. Identified Risks (with evidence)
3. Edge Case Gaps
4. Required Actions (P0 / P1 / P2)
5. Suggested Fix Snippets (if applicable)

Procedure

Gate 0 – Change Mapping

Break the diff into logical Change Units.

For each Change Unit, summarize:

• Before behavior
• After behavior
• Intended purpose

Gate 1 – Side Effect Analysis

> EFFICIENCY RULE: Before deep-analyzing each dimension, perform a > quick relevance check based on the Change Units from Gate 0. If a > dimension has zero intersection with the change scope (e.g., no > database writes exist in any Change Unit → State mutations is > not relevant), note "N/A — no relevant changes" and move to the next > dimension. Do NOT skip the relevance check itself. > "Zero intersection" means the Change Units contain no code that could > possibly affect that dimension. If uncertain or indirect, treat as relevant.

For each Change Unit, explicitly evaluate:

• Public contract changes
• Input schema
• Output structure
• Error types
• HTTP status codes
• Ordering or pagination behavior

• State mutations
• Database writes
• Transaction boundaries
• Cache keys
• Idempotency guarantees
• Retry logic

• Concurrency & ordering
• Async behavior
• Race conditions
• Locking
• Event ordering

• Resource usage
• N+1 queries
• Increased memory footprint
• Large payload buffering
• CPU-heavy loops

• Security implications
• Authorization checks
• Input validation
• Injection vectors
• Secret or PII leakage in logs

• Observability impact
• Log structure changes
• Metric cardinality explosion
• Missing tracing

• Operational risk
• Rollback safety
• Migration requirements
• Feature flag need

Each risk must include:

• Evidence (file:line reference)
• Impact assessment
• Mitigation proposal

Gate 2 – Edge Case Matrix

> SCOPING RULE: Apply edge case verification only to code paths > that were modified or newly introduced in the current diff. > Pre-existing edge case gaps in unchanged code are out of scope > for this audit.

Explicitly verify handling of:

• Null / empty values
• Boundary values (min/max/overflow)
• Large inputs
• Partial failures (timeouts, downstream 5xx)
• Duplicate or out-of-order events
• Permission boundary violations
• Backward compatibility

For each:

• Expected behavior
• Actual behavior
• Gap analysis
• Fix proposal

Gate 3 – Regression Proof

Identify:

• Covered scenarios (existing tests)
• Uncovered risk areas

Propose minimal additional tests (1–3 max) for high-risk gaps.

If adding tests is not feasible:

• Add guard clauses
• Add assertions
• Improve logging
• Provide safe fallback

Gate 4 – Code Hygiene Check

Verify:

• No dead code
• No unnecessary abstraction
• No silent behavioral drift
• No inconsistent error handling
• No policy values hardcoded without documentation

Guardrails

• Do not invent missing facts. Every risk must cite evidence (file:line).
• Explicitly state assumptions when context is incomplete.
• If context is insufficient to assess a risk area, ask for clarification.
• Respect constraints and non-goals stated in the change description.
• Do not inflate severity — use P0/P1/P2 classification honestly.
• Do not skip risk dimensions because they seem unlikely — a dimension may be

marked N/A only through the Gate 1 relevance check, never on intuition.

Failure Patterns

Common bad outputs:

• Reviewing only the diff without considering the broader call chain
• Classifying all findings as P0 (no prioritization)
• Asserting risks without file:line evidence
• Skipping edge case matrix entirely
• Producing generic advice instead of change-specific findings
• Missing concurrency or state mutation risks in async code paths

Example 1 (Minimal Context)

Input:

Single function change: getUserById in services/user.ts now returns null instead of throwing NotFoundError when user is not found.

Output:

1. Change Units: 1 unit — getUserById return type change
2. Identified Risks:

• P0: All callers expecting NotFoundError will now receive null without handling it (routes/user.ts:45, middleware/auth.ts:23)
• P1: API response may change from 404 to 200 with null body if caller does not check

1. Edge Case Gaps: caller with try/catch only — null passes through silently
2. Required Actions: P0 — update all 3 callers to handle null return; P1 — add null-check test for each caller
3. Suggested Fix: add if (!user) throw new NotFoundError() guard in each route handler

Example 2 (Realistic Scenario)

Input:

8 files changed: new caching layer added to services/product.ts, cache invalidation on PUT /products/:id, new Redis dependency in config/cache.ts. Existing tests cover CRUD but not cache behavior.

Output:

1. Change Units: 3 units — (a) cache read in getProduct, (b) cache invalidation on update, (c) Redis config setup
2. Identified Risks:

• P0: Cache key uses product:${id} but bulk endpoints use different ID format — stale cache risk (services/product.ts:78)
• P1: No TTL set — cache entries persist indefinitely (config/cache.ts:12)
• P1: Redis connection failure has no fallback — entire product API will 500 (services/product.ts:34)
• P2: Cache invalidation only on PUT, not on DELETE (routes/product.ts:92)

1. Edge Case Gaps: concurrent PUT requests may cause cache race condition; no handling for Redis timeout
2. Required Actions: P0 — normalize cache key format; P1 — add TTL and Redis fallback; P1 — add DELETE invalidation; P2 — add cache-miss metric
3. Suggested Fix: wrap Redis calls in try/catch with DB fallback, add EX 3600 to SET commands

Notes

FAST MODE (only if explicitly requested):

• Limit to top 5 likely risks
• Require at least one regression safeguard