ADR-007: Input Validation for Injection Prevention

Status: Accepted Date: 2026-01-12 PR: #75 Issue: #48 Milestone: v0.2.0

Context

DynamoDB keys use # as a delimiter (e.g., ENTITY#user-1#BUCKET#gpt-4). Without validation, malicious input containing # could:

• Traverse to different record types
• Access other entities' data
• Corrupt key structure

This is analogous to SQL injection but for DynamoDB key-based access patterns.

Decision

Add comprehensive input validation at model construction time, forbidding the # delimiter character in all user-provided identifiers.

Validation rules:

Field	Pattern	Max Length
entity_id, parent_id	^[a-zA-Z0-9][a-zA-Z0-9_.\-:@]*$	256
limit_name, resource	^[a-zA-Z][a-zA-Z0-9_.\-]*$	64

New exceptions: - ValidationError (base) - InvalidIdentifierError (entity_id, parent_id) - InvalidNameError (limit_name, resource)

Consequences

Positive: - Injection attacks impossible by construction - Early failure with clear error messages - Supports common ID formats: UUIDs, API keys, emails - Validation happens once at model creation, not on every operation

Negative: - Some previously-valid IDs may be rejected (containing #) - Migration required for existing data with invalid characters

Alternatives Considered

• Escape/encode special characters: Rejected; complex, error-prone, obscures data in DynamoDB console
• Validate only at repository layer: Rejected; too late, models could be passed around with invalid data
• Allowlist specific formats (UUID only): Rejected; too restrictive for diverse use cases