Scanning Pipeline

PromptShield scans text using a deterministic detector pipeline.

Each detector identifies a specific class of prompt manipulation technique.

The pipeline is designed to run quickly and produce stable results.

Pipeline execution model

The scanner executes detectors sequentially:

for (const detector of detectors) {
  threats.push(...detector(text, options, context));
}

Detectors do not modify text.

They only report threats.

Context sharing

Detectors share a mutable ScanContext.

This allows:

• reuse of line offsets
• location mapping
• performance optimizations

Example:

context.lineOffsets

This prevents recomputation across detectors.

stopOnFirstThreat option

The scanner supports early termination:

stopOnFirstThreat: true

This is useful for:

• CI validation
• large document scanning
• performance-sensitive environments

Severity filtering

The scanner supports severity filtering:

minSeverity

Detectors respect this option to skip lower-severity checks.

Example:

minSeverity: "HIGH"

Deterministic behavior

The scanning pipeline guarantees:

• same input → same output
• no randomness
• no AI models
• no network calls
• stable reporting order

This is critical for CI usage.

Performance considerations

The pipeline is optimized for:

• editor typing latency
• incremental validation
• workspace scanning
• large file handling

All detectors operate in linear time relative to text size.

Extending the pipeline

New detectors should:

• be deterministic
• follow the ThreatReport contract
• avoid modifying text
• respect ScanOptions
• be placed according to dependency order

Injection detection should remain the final stage.